Glossary

238 terms used across the playgrounds, in plain language. Each entry links to the pages that use it.

A

• Abduction: Moving a limb away from the body’s midline (positive roll on Marvin Bot). On: 01 Degrees of freedom and kinematic chains
• Accelerometer: A sensor that measures gravity’s pull plus the body’s own acceleration, so at rest it tells which way is down. On: 11 Sensors and compute
• Action: What the policy outputs at each step. For Marvin Bot, one number per joint: a target angle offset from the standing pose, which a fast PD loop in each motor then tries to reach. On: 15 PD control, gains and armature, 16 Reinforcement learning for walking
• Action chunk: A short sequence of future actions output at once by a policy, executed while the next chunk is computed. On: 18 Whole-body control and VLA models
• Action scale: The factor converting the policy’s output into a joint-angle offset from the stance; 0.5 rad per unit for Marvin Bot. On: 16 Reinforcement learning for walking
• Active stereo: Stereo depth with an infrared projector adding a dot pattern, so that textureless surfaces such as plain floors can still be matched. On: 11 Sensors and compute
• Actuator: A motor with its gearbox, encoder and driver electronics in one unit. On: 03 System architecture: from battery to policy
• Actuator bench: A test rig for single actuators. It measures torque, speed, heating and communication timing before many units are bought. On: 22 Build phases and roadmap
• Actuator delay: Time between a command being computed and the motor acting on it; Marvin Bot’s training randomises 0-10 ms. On: 15 PD control, gains and armature
• Actuator network: A small neural network trained on bench data to predict an actuator’s torque from recent commands and motion, used in simulation instead of an ideal motor model. On: 20 Sim-to-real and system identification
• Adaptive sampling: Starting training episodes more often at the parts of the clips where the policy has been failing. On: 17 Motion tracking: Kimodo, retargeting and BeyondMimic
• Advantage: During training: how much better an action turned out than expected. A second network (the critic) predicts how well things should go from a situation; if the action did better than that, it gets a positive advantage and becomes more likely next time. On: 16 Reinforcement learning for walking
• AHRS: Attitude and heading reference system: an IMU (motion sensor) with its own built-in processing, so it reports the body’s tilt and heading directly instead of raw rotation rates and accelerations. On: 11 Sensors and compute
• Aliasing: When a signal is sampled too slowly, fast vibrations masquerade as slow ones; sampling fast and averaging prevents it. On: 11 Sensors and compute
• Anchor body: The body (usually the torso) whose pose is tracked globally; the other bodies are compared relative to it, so drift in position is not counted as pose error. On: 17 Motion tracking: Kimodo, retargeting and BeyondMimic
• Ankle strategy: Balancing by shifting pressure between heel and toes with ankle torque, without stepping. On: 02 Balance: centre of mass, support polygon, ZMP and capture point
• Anthropometry: The measurement of human body dimensions. Tables give segment lengths as fractions of standing height. On: 07 Mass budget and proportions
• Arbitration: How CAN decides who talks when two devices start at once: each sends its identifier number bit by bit, and the device with the lower number wins and carries on, while the other waits. Nothing is garbled. On: 10 CAN buses and bandwidth
• Armature: The motor rotor’s inertia as felt at the joint (the reflected inertia: rotor inertia times the gear ratio squared). Simulators add it to the joint’s own inertia; Marvin Bot’s values are placeholders until the phase 2 bench measures them. On: 15 PD control, gains and armature, 20 Sim-to-real and system identification
• Asymmetric actor-critic: A training trick: the policy that will run on the robot (the actor) sees only what real sensors can measure, while the helper network that grades it during training (the critic) is allowed to see simulator-only facts such as the true speed or the exact terrain. On: 16 Reinforcement learning for walking

B

• Back-EMF: The voltage a spinning motor generates against its supply, proportional to speed. It sets the top speed for a given battery voltage. On: 04 Actuators: quasi-direct drive, torque, speed and heat
• Backdrivability: How easily the output can be turned by an external force. Backdrivable actuators absorb impacts and can sense torque from motor current. On: 04 Actuators: quasi-direct drive, torque, speed and heat
• Baseline: The distance between a stereo camera’s two cameras; a longer baseline gives smaller depth error at range. On: 11 Sensors and compute
• Bending moment: The turning effect of a load about a section of a beam: force times distance, in N·m. On: 08 Structure and code-CAD
• Bias: A slowly drifting offset in a sensor’s reading; integrated gyro bias makes an angle estimate drift without limit. On: 11 Sensors and compute
• Bill of materials (BOM): The complete list of parts, quantities and prices needed to build a product. Marvin Bot’s is the spreadsheet bom/bom.xlsx. On: 21 Bill of materials and costs
• Bit stuffing: A CAN rule: after five identical bits in a row, the sender inserts one opposite bit so receivers can keep their clocks in step. It means a frame’s length depends slightly on the data it carries. On: 10 CAN buses and bandwidth
• BMS: Battery management system: electronics that watch every cell’s voltage and temperature, balance them and disconnect the pack on a fault. On: 09 Power and battery
• Bought part: A part bought finished, such as an actuator, camera or bearing. Its cost is the price on a listing. On: 21 Bill of materials and costs
• Bus load: The fraction of time a bus is busy transmitting frames. On: 10 CAN buses and bandwidth

C

• C-rate: Current divided by a cell’s capacity: 1C empties it in an hour, 10C in six minutes. On: 09 Power and battery
• CAN bus: Controller Area Network: a two-wire differential bus on which every node hears every frame and identifiers set priority; RobStride actuators use classic CAN at 1 Mbit/s. On: 03 System architecture: from battery to policy, 10 CAN buses and bandwidth
• CAN FD: CAN with flexible data rate: the data phase runs faster than arbitration and frames carry up to 64 bytes. On: 10 CAN buses and bandwidth
• Capture point: The spot on the ground where the robot would have to step (or put its centre of pressure) to come to a complete stop. The faster the body is moving, the further ahead that spot is. On: 02 Balance: centre of mass, support polygon, ZMP and capture point
• Carrier board: The board a compute module such as a Jetson plugs into. It provides the connectors and power. A developer kit is a module already on a carrier board. On: 07 Mass budget and proportions, 21 Bill of materials and costs
• Centre of mass: The mass-weighted average position of all a robot’s parts; gravity acts as if all the weight were there. Balance depends on keeping it over the feet. On: 02 Balance: centre of mass, support polygon, ZMP and capture point, 07 Mass budget and proportions
• Centre of pressure: The single point where the ground’s distributed push on the soles can be treated as acting. On: 02 Balance: centre of mass, support polygon, ZMP and capture point
• Chirp signal: A test input whose frequency sweeps from low to high, used to excite a system across its frequency range. On: 20 Sim-to-real and system identification
• Class-T fuse: A fast, current-limiting fuse with a high DC interrupting rating, used to break a hard short of a lithium pack safely. On: 09 Power and battery
• Closed kinematic chain: A set of links and joints that forms a loop, so the joint angles are not independent; most simulators need an extra constraint to handle it. On: 06 The parallel-linkage ankle
• Code-CAD: Writing mechanical parts as programs that build solids from primitives and boolean operations, so they are parametric, diffable and can feed other tools directly. On: 08 Structure and code-CAD
• Coefficient of friction: μ, the ratio of the largest sideways friction force to the pressing force at a contact; soft rubber has a higher μ than hard plastic. On: 19 Hands
• Coefficient of restitution: Ratio of rebound speed to impact speed; 0 means no bounce, 1 a perfectly elastic bounce. On: 14 Simulation: MuJoCo, Isaac Lab, Newton and PhysX
• Collision geometry: Simple shapes the physics engine uses to detect contact; kept simple because contact checks run every step. On: 13 The robot description: one source of truth
• Collision sweep: Moving each joint through its range in software and checking for overlaps between parts, to set joint limits and find packaging errors. On: 08 Structure and code-CAD
• Complementary filter: A simple way to estimate tilt from an IMU: trust the gyroscope for quick changes (it is smooth but slowly drifts) and the accelerometer’s sense of “down” for the long-term average (it is noisy but does not drift). On: 11 Sensors and compute
• Constraint solver: The part of a physics engine that works out all the contact and joint forces at once each step, so that feet do not sink into the floor and joints stay connected. PGS, TGS and MuJoCo’s Newton solver are different methods for this. On: 14 Simulation: MuJoCo, Isaac Lab, Newton and PhysX
• Contactor: A heavy-duty relay that switches the main battery current; on Marvin Bot the e-stop chain opens it. On: 09 Power and battery, 12 Safety
• Control hierarchy: Stacking control loops at different rates, each slower loop setting targets for a faster one below it: VLA at a few hertz, policy at 50 Hz, joint PD and current loops at kilohertz. On: 03 System architecture: from battery to policy
• Copper loss: Heat made by current flowing through the motor’s copper windings. It grows with the current squared, so doubling the torque roughly quadruples the heat. On: 04 Actuators: quasi-direct drive, torque, speed and heat
• Coulomb friction: A constant resisting torque that opposes motion regardless of speed, from gear teeth and seals. On: 20 Sim-to-real and system identification
• Coupled joint limit: A joint limit that depends on another joint’s angle; on Marvin Bot the shoulder roll may go below 0 degrees only with the arm swung forward. On: 08 Structure and code-CAD, 13 The robot description: one source of truth, 17 Motion tracking: Kimodo, retargeting and BeyondMimic
• Crank: A short arm fixed to a motor shaft; its tip moves on a circle and pushes or pulls a rod. On: 06 The parallel-linkage ankle
• Curriculum: Training that starts with easy tasks (flat ground, slow speeds) and makes them harder as the policy improves. On: 16 Reinforcement learning for walking
• Cycloidal drive: A gearbox in which a lobed disc rolls inside a ring of pins, sharing load over many lobes; compact and shock-tolerant. On: 04 Actuators: quasi-direct drive, torque, speed and heat

D

• Damped least squares: A way to solve inverse kinematics step by step: nudge every joint in the direction that moves the foot (or hand) towards its target, with a little damping so the steps stay small and stable near awkward poses where the leg is straight. On: 00 Marvin Bot in 3-D
• Damping (kd): The “D” setting of a PD controller (in N·m·s/rad): how strongly a joint pushes back against moving fast, like a door closer. More damping means less wobble but slower motion. On: 15 PD control, gains and armature
• Damping mode: All joints commanded to resist motion like dampers, with no position target, so a falling robot crumples instead of thrashing. On: 12 Safety
• Damping ratio: A single number that says how a spring-like joint settles: below 1 it overshoots and wobbles, at exactly 1 it arrives as fast as possible without overshooting, above 1 it creeps in slowly. Computed as kd / (2 sqrt(kp J)). On: 15 PD control, gains and armature
• DC-DC converter: An electronic circuit that converts one DC voltage to another, such as 48 V to 12 V. On: 09 Power and battery
• Decoupled control: Splitting a humanoid’s control into a learned lower-body balance/locomotion policy and a separate arm controller such as IK. On: 18 Whole-body control and VLA models
• Defence in depth: Several independent safety layers, so a fault one layer misses is caught by the next. On: 12 Safety
• Degree of freedom: One independent number needed to describe a configuration, such as one joint angle. A rigid body in space has 6. On: 01 Degrees of freedom and kinematic chains
• Depth of discharge: The share of a battery’s capacity that is actually used between charges. On: 09 Power and battery
• Developer kit: A ready-to-use board sold for development, with the module, carrier board, cooling and ports. Marvin Bot uses the Jetson AGX Orin 64 GB developer kit. On: 21 Bill of materials and costs
• Diffusion model: A generative neural network that produces data (images, motions, actions) by gradually removing noise from a random start, guided by a condition such as a text prompt. On: 17 Motion tracking: Kimodo, retargeting and BeyondMimic
• Discount factor: A number γ between 0 and 1 that sets how much future rewards count; 0.99 at 50 Hz means the policy looks about 2 s ahead. On: 16 Reinforcement learning for walking
• Disparity: How far a point shifts between a stereo camera’s two images, in pixels; depth is focal length times baseline divided by disparity. On: 11 Sensors and compute
• Distal mass: Mass far from the joint that swings a limb, such as motors at the ankle. It costs inertia as the square of its distance. On: 06 The parallel-linkage ankle, 07 Mass budget and proportions
• Divergent component of motion: Another name for the capture point used in walking control: the part of the body’s motion that keeps running away (like a falling broom) unless the feet push in the right place. On: 02 Balance: centre of mass, support polygon, ZMP and capture point
• Domain randomisation: Varying physical parameters such as friction, mass and delays across simulated episodes so the policy works over a range that includes the real robot. On: 16 Reinforcement learning for walking, 20 Sim-to-real and system identification
• Dorsiflexion: Bending the ankle so the toes move up towards the shin (positive ankle pitch on Marvin Bot). On: 01 Degrees of freedom and kinematic chains
• Duty cycle: The share of a cycle during which a load is applied, here the stance share of a step. On: 05 Leg loads and stairs

E

• E-stop: Emergency stop: a latching control that removes hazardous power; on Marvin Bot it opens the main contactor. On: 12 Safety
• Effective mass: How heavy a moving part of the robot feels to whatever it hits. A long arm swung from the shoulder can feel lighter or heavier at the hand than its real mass, depending on how its mass is spread out. On: 12 Safety
• Encoder: A sensor that measures a joint’s angle; RobStride actuators use absolute magnetic encoders. On: 11 Sensors and compute
• Episode: One run of a simulated robot from its start until it falls, is reset or times out. Training averages over many thousands of episodes running in parallel. On: 16 Reinforcement learning for walking
• Execution horizon: How many actions of each chunk are executed before switching to the next chunk. On: 18 Whole-body control and VLA models
• Exploded view: A picture of an assembly with its pieces pulled apart along the directions they fit together, so you can see which part attaches to which. On: 00 Marvin Bot in 3-D
• Extended identifier: The longer, 29-bit form of a CAN message’s identifier, the label that says what a message is and sets its priority (the standard form has 11 bits). Marvin Bot’s RobStride motors use the 29-bit form. On: 10 CAN buses and bandwidth

F

• Far-side bearing: A second bearing on the opposite side of a joint from the motor output, so the link is supported on both sides and side loads do not bend thin plates. On: 08 Structure and code-CAD
• Field weakening: A motor-control technique that partly cancels the magnets’ field to run above the normal top speed, at the cost of torque and extra heat. On: 04 Actuators: quasi-direct drive, torque, speed and heat
• Field-oriented control: The method a brushless motor driver uses to steer current through the motor’s three coils thousands of times a second, so the motor produces exactly the torque that was asked for. On: 03 System architecture: from battery to policy
• Floating base: A robot body, like a humanoid’s pelvis, that is not bolted down; its 6 degrees of freedom are not driven by any motor. On: 01 Degrees of freedom and kinematic chains
• Flow matching: A way of training a model to produce data (here, short sequences of robot motion) by learning how to nudge random noise, step by step, towards a realistic sample. It needs only a few steps, so it is fast enough for robots. On: 18 Whole-body control and VLA models
• Foot sliding (skate): Horizontal movement of a foot that should be planted, a common artefact of generated or retargeted motion, measured in cm/s. On: 17 Motion tracking: Kimodo, retargeting and BeyondMimic
• Forward kinematics: Computing where each link of the robot is, and how it is oriented, from the joint angles and segment lengths, working outwards joint by joint from the pelvis. On: 01 Degrees of freedom and kinematic chains, 13 The robot description: one source of truth, 17 Motion tracking: Kimodo, retargeting and BeyondMimic
• Frame: One CAN message: identifier, control bits, up to 8 data bytes (64 in CAN FD), checksum and acknowledgement. On: 10 CAN buses and bandwidth
• Friction cone: The range of directions a foot (or finger) can push in without slipping: the sideways force can be at most μ times the pressing force, which traces out a cone around the surface normal. On: 14 Simulation: MuJoCo, Isaac Lab, Newton and PhysX
• Full cost to replicate: What someone starting from nothing would pay for all parts, with owned items still counted at their current price. On: 21 Bill of materials and costs

G

• GAE: Generalised Advantage Estimation: a recipe used in PPO training for scoring how good each action was, by blending short-term evidence (less noisy but more biased) with long-term evidence (more noisy). A parameter λ sets the mix. On: 16 Reinforcement learning for walking
• Gantry: A frame with a harness above the robot. It catches the robot before it hits the floor during early standing and walking tests. On: 12 Safety, 22 Build phases and roadmap
• Gear ratio: How many turns the motor makes for one turn of the output. A ratio of 9 (RS04, RS03) multiplies torque by about 9 and divides speed by 9; it also multiplies the rotor’s inertia felt at the joint by 81 (the ratio squared). The RS10P’s 25:1 gear gives a small motor 42 N·m, at a 13 rad/s top speed. On: 04 Actuators: quasi-direct drive, torque, speed and heat, 15 PD control, gains and armature
• Gear stage: An extra reduction between motor and joint: multiplies torque by the ratio and divides speed by it. On: 05 Leg loads and stairs
• gs_usb: An open standard for USB-to-CAN adapters (used by candleLight adapters); Linux has a driver for it built in, so no extra software is needed. On: 10 CAN buses and bandwidth
• Gyroscope: A sensor that measures angular velocity (how fast something rotates) about three axes. On: 11 Sensors and compute

H

• Harmonic drive: A strain-wave gearbox giving high ratios (30:1 to 160:1) with no backlash, but hard to backdrive and sensitive to shock loads. On: 04 Actuators: quasi-direct drive, torque, speed and heat

I

• Ideal-diode OR: A circuit that lets two power sources feed the same device, so if one is removed the other takes over instantly, without either one pushing current back into the other. It uses transistors that act like one-way valves with almost no loss. On: 09 Power and battery
• Impulse: Force times the time it acts (N·s); a short push of impulse J changes the CoM velocity by J divided by the mass. On: 02 Balance: centre of mass, support polygon, ZMP and capture point
• IMU: Inertial measurement unit: a gyroscope and an accelerometer that measure rotation rate and the direction of gravity. On: 03 System architecture: from battery to policy, 11 Sensors and compute
• Inertia tensor: Nine numbers (a 3 x 3 table, in kg·m²) that describe how hard a part is to spin about each direction through its centre of mass. A long thin rod, for example, spins easily about its length and hard end over end. On: 13 The robot description: one source of truth
• Integrator: The rule that turns accelerations into new velocities and positions each step, such as semi-implicit Euler or Runge-Kutta 4. On: 14 Simulation: MuJoCo, Isaac Lab, Newton and PhysX
• Internal resistance: The small resistance inside every cell that makes its voltage sag under load and rise under charge. On: 09 Power and battery
• Inverse kinematics: The reverse problem: finding joint angles that put a link at a desired position and orientation. On: 01 Degrees of freedom and kinematic chains, 18 Whole-body control and VLA models

J

• Jacobian: A table of numbers that says how fast a point such as the hand or foot moves when each joint turns a little. Used the other way round, it converts a force at the hand or foot into the torque each joint must supply. On: 05 Leg loads and stairs, 06 The parallel-linkage ankle, 18 Whole-body control and VLA models
• Jitter: Variation in timing from one cycle to the next. On: 10 CAN buses and bandwidth
• Joint: A connection between a parent and a child link that allows motion, here a hinge about one axis with limits. On: 13 The robot description: one source of truth
• Joint limit (range): The smallest and largest angle a joint is allowed to reach, set by where parts would collide or a linkage runs out of travel. Marvin Bot’s ranges are written once in description/robot.yaml. On: 01 Degrees of freedom and kinematic chains, 13 The robot description: one source of truth

K

• Keep-alive battery: A small battery that powers only the computer so it stays on while the main pack is swapped. On: 09 Power and battery
• Keypoint: A tracked point on a body, such as a hip, knee, wrist or shoulder, used to compare two bodies’ poses. On: 17 Motion tracking: Kimodo, retargeting and BeyondMimic
• Kinematic chain: Links connected by joints. In a serial chain such as a leg, each link is carried by the one before it. On: 01 Degrees of freedom and kinematic chains

L

• Landed cost: The price of a part once shipping, import duty and taxes are added. The Marvin Bot BOM does not include these. On: 21 Bill of materials and costs
• Latency: The delay between something happening (a tilt, a command) and the system acting on it; it adds up stage by stage through sensors, computer, bus and driver. On: 03 System architecture: from battery to policy, 20 Sim-to-real and system identification
• Lead time: The time between ordering a part and having it in hand. It does not shrink however many hours you work. On: 22 Build phases and roadmap
• Least squares: Choosing model parameters that minimise the sum of squared differences between the model’s predictions and the measurements. On: 20 Sim-to-real and system identification
• Lever arm: The perpendicular distance from a joint axis to the line along which a force acts. On: 05 Leg loads and stairs
• Linear inverted pendulum: The simplest model of a walking robot: all the mass in one point at a fixed height, balanced on a weightless stick leg, like a broom balanced on your hand. On: 02 Balance: centre of mass, support polygon, ZMP and capture point
• Link: One rigid body of a robot model, such as a thigh or a foot. On: 13 The robot description: one source of truth
• Linkage drive: Moving finger joints through rigid bars from motors at the finger base or in the palm. On: 19 Hands
• Load path: The chain of parts a force travels through, for example from the foot to the pelvis during a landing. On: 08 Structure and code-CAD

M

• Made part: A part designed in the project and produced from raw material, by CNC machining or 3D printing. The BOM prices only its inputs: stock, filament, tools or a machining service. On: 21 Bill of materials and costs
• Mass budget: A list of every part of a machine with its mass and how that mass is known (estimate, CAD, datasheet, weighed), summed and checked against a target and a limit. On: 07 Mass budget and proportions
• Massively parallel simulation: Running thousands of copies of the robot at once on one GPU to collect reinforcement-learning experience quickly. On: 14 Simulation: MuJoCo, Isaac Lab, Newton and PhysX
• Mirror rule: How the right side is built from the left: y coordinates negated, roll and yaw axes negated, so equal angles mean the same motion on both sides. On: 13 The robot description: one source of truth
• MIT mode: A way of commanding a motor in which each message gives a target angle and speed plus how stiff and how damped the joint should be (kp and kd) and an extra torque; the motor driver then works out the torque itself, very fast. On: 04 Actuators: quasi-direct drive, torque, speed and heat, 10 CAN buses and bandwidth
• MJCF: MuJoCo’s own robot file format (XML). It holds what a URDF holds plus simulator settings such as motor models, joint friction and contact behaviour. On: 13 The robot description: one source of truth
• Module (assembly): A section of the robot that is built and tested on its own and then joined to the rest at a few bolted flanges and connectors: Marvin Bot has seven (torso core, pelvis, two legs, two arms, head), plus the battery and the optional shells. On: 00 Marvin Bot in 3-D
• Moment arm: The perpendicular distance from a joint axis to the line along which a force acts (for a weight, the horizontal distance); torque equals force times moment arm. On: 19 Hands
• Moment of inertia: A body’s resistance to angular acceleration about an axis: the sum of each mass times its distance from the axis squared, in kg·m². On: 06 The parallel-linkage ankle, 07 Mass budget and proportions
• Motion tracking: Training a policy to follow a reference motion clip, rewarded for staying close to it. A new skill is then a new clip rather than new reward code. On: 17 Motion tracking: Kimodo, retargeting and BeyondMimic

N

• Natural frequency: How fast a joint held by a PD controller would naturally swing back and forth, like a mass on a spring: stiffer settings or lighter limbs give a higher frequency. Computed as sqrt(kp / J) in rad/s (divide by 2π for Hz). On: 15 PD control, gains and armature
• Neutral axis: In a bent beam, the line where the material is neither stretched nor squeezed: on one side it stretches, on the other it compresses. On: 08 Structure and code-CAD
• No-load speed: The fastest an actuator turns with no load at its rated voltage, where back-EMF uses up the whole supply voltage. On: 04 Actuators: quasi-direct drive, torque, speed and heat
• Noise density: How noisy a sensor is, given in a form that does not depend on how fast you read it. The total noise you see grows with the square root of the bandwidth (roughly, how fast the readings are allowed to change). On: 11 Sensors and compute
• Null space: The joint motions that leave a task untouched. With your hand resting on a table, you can still move your elbow: that elbow motion is in the null space of the hand task, and a controller can use it for a lower-priority goal. On: 18 Whole-body control and VLA models

O

• Observation: Everything the policy is shown at each step: for Marvin Bot’s walking policy, the body’s angular velocity, which way gravity points, the speed command, the joint angles and speeds, and its own previous action. On: 03 System architecture: from battery to policy, 16 Reinforcement learning for walking
• On-policy: A kind of learning that only learns from experience collected by the current version of the policy; once the policy is updated, the old experience is thrown away and fresh experience is collected. On: 16 Reinforcement learning for walking
• ONNX: Open Neural Network Exchange: a portable file format for trained networks, used to move a policy from the training code to the robot. On: 03 System architecture: from battery to policy
• Overshoot: How far a step response goes past its target, as a percentage of the step. On: 15 PD control, gains and armature

P

• Parallel mechanism: A mechanism in which several actuators drive the same output through separate paths, so each load is shared. On: 06 The parallel-linkage ankle
• Pass check: A test at the end of a build phase. It decides whether the project moves on or fixes something first. On: 22 Build phases and roadmap
• PD control: Proportional-derivative control, the simplest joint controller: the motor pushes towards the target angle with a strength kp (like a spring) and resists speed with a strength kd (like a damper). Torque = kp × (target − angle) − kd × speed. On: 15 PD control, gains and armature
• Peak torque: The most torque an actuator can give in short bursts, limited by current; it overheats in seconds if held. On: 04 Actuators: quasi-direct drive, torque, speed and heat, 05 Leg loads and stairs
• Penalty contact: Contact modelled as a stiff spring and damper that pushes back in proportion to how far bodies overlap. On: 14 Simulation: MuJoCo, Isaac Lab, Newton and PhysX
• Physics engine: Software that steps a model forward in time: detects contacts, computes constraint forces and integrates motion. On: 14 Simulation: MuJoCo, Isaac Lab, Newton and PhysX
• Pinch grasp: Holding an object between fingertip and thumb, where friction at the two contacts must support its weight. On: 19 Hands
• Pocketing: Milling recesses into a plate down to a thin floor, leaving full-depth ribs, to save mass while keeping bending strength. On: 08 Structure and code-CAD
• Policy: The trained neural-network controller: every control tick (50 Hz on Marvin Bot) it maps what the robot observes to the actions it takes, here joint position targets. On: 02 Balance: centre of mass, support polygon, ZMP and capture point, 03 System architecture: from battery to policy, 16 Reinforcement learning for walking
• Power grasp: Holding an object wrapped in the curled fingers and palm, so geometry carries the load rather than fingertip friction alone. On: 19 Hands
• PPO: Proximal Policy Optimisation: the trial-and-error learning method used to train Marvin Bot’s walking. Its key idea is to change the policy only a little at each update, so training stays stable instead of lurching between behaviours. On: 16 Reinforcement learning for walking
• Precharge: Filling the motor drivers’ input capacitors slowly through a resistor before the main contactor closes, to avoid a huge inrush current. On: 09 Power and battery
• Price basis: A note saying where a price came from and when: a dated listing, an unverified reseller page or an estimate. On: 21 Bill of materials and costs
• Price per N·m: Actuator price divided by its peak or continuous torque. It gives a first comparison of value across sizes and brands. On: 21 Bill of materials and costs
• Projected gravity: The gravity direction expressed in the pelvis frame; the policy uses it instead of tilt angles. On: 11 Sensors and compute
• Proprioceptive actuation: Controlling contact forces using only the actuator’s own current and position sensing, without force sensors; enabled by low gear ratios. On: 04 Actuators: quasi-direct drive, torque, speed and heat
• Push rod: A rigid rod of fixed length that transmits force between a crank and the driven part through joints at each end. On: 06 The parallel-linkage ankle
• Push-off: The moment at the end of a step when the back foot lifts its heel and pushes through the toes before leaving the ground. On a stair it is when the leading leg starts carrying the whole body. On: 00 Marvin Bot in 3-D

Q

• Quadratic program (QP): A kind of optimisation problem that computers solve very quickly: minimise a sum of squared errors while respecting limits written as simple linear rules. Classical whole-body controllers solve one every control cycle. On: 18 Whole-body control and VLA models
• Quantity tier: A lower unit price for ordering at least a set number of one item, such as Seeed’s 10+ price for RobStride actuators. On: 21 Bill of materials and costs
• Quasi-direct drive: An actuator with a large-diameter motor and a single low-ratio gear stage (about 6:1 to 10:1), keeping low reflected inertia and backdrivability while giving enough torque. On: 03 System architecture: from battery to policy, 04 Actuators: quasi-direct drive, torque, speed and heat
• Quasi-static: Treating a movement as a series of still poses, so only gravity (and the floor’s push) is counted and accelerations are ignored. A good first estimate for slow movements; real fast movements need more torque. On: 00 Marvin Bot in 3-D
• Quasi-static contact: Contact in which a body part is clamped and pressed, as opposed to a brief transient impact. On: 12 Safety

R

• Rated (continuous) torque: The torque an actuator can hold indefinitely without its winding passing the temperature limit, measured with a specified heat sink. On: 04 Actuators: quasi-direct drive, torque, speed and heat, 05 Leg loads and stairs, 07 Mass budget and proportions, 19 Hands
• Reduction stage: An extra gear, belt or linkage between motor and joint that multiplies torque and divides speed by its ratio. On: 07 Mass budget and proportions
• Reference motion: The recorded or generated motion a tracker should follow: where each body part is at every moment. On: 17 Motion tracking: Kimodo, retargeting and BeyondMimic
• Reflected inertia: The motor rotor’s inertia as felt at the joint: N² times the rotor inertia for a gear ratio N. It adds to impact forces. On: 04 Actuators: quasi-direct drive, torque, speed and heat, 05 Leg loads and stairs, 15 PD control, gains and armature
• Regenerative braking: A motor slowing a moving load acts as a generator and pushes energy back onto the power bus. On: 09 Power and battery
• Reinforcement learning: Learning a behaviour by trial and error: an agent acts, receives a numerical reward, and adjusts itself to collect more reward over time. On: 16 Reinforcement learning for walking
• Retargeting: Converting a motion from one body (usually a person) to another (the robot) with different sizes, joints and limits, typically by optimising the robot’s joint angles to match scaled keypoints. On: 17 Motion tracking: Kimodo, retargeting and BeyondMimic
• Return: The sum of future rewards from a moment on, with later rewards counted less by the discount factor. On: 16 Reinforcement learning for walking
• Revolute joint: A hinge that allows rotation about one axis and nothing else: one degree of freedom. On: 01 Degrees of freedom and kinematic chains
• Reward: A number the training environment gives the policy after every step to score what it just did (for example, positive for following the commanded speed, negative for wasting torque). Reinforcement learning adjusts the policy to collect as much total reward as possible. On: 16 Reinforcement learning for walking, 17 Motion tracking: Kimodo, retargeting and BeyondMimic
• Reward hacking: When an optimiser finds behaviour that scores well on the reward but is not what the designer intended, such as crouching or standing still. On: 16 Reinforcement learning for walking
• Reward kernel: A formula that turns an error (how far off the robot is) into a reward between 0 and 1: exactly 1 when the error is zero, about 0.37 when the error equals a chosen tolerance σ, and almost 0 beyond twice that tolerance. On: 17 Motion tracking: Kimodo, retargeting and BeyondMimic
• Riser: The vertical height of one stair step. On: 05 Leg loads and stairs
• Risk register: A list of things that could be wrong, each with how and when it will be checked. On: 22 Build phases and roadmap
• Risk-driven development: Planning work so the biggest remaining uncertainty is tested first, with the cheapest build that can expose it. On: 22 Build phases and roadmap
• RMS torque: Root-mean-square torque over a cycle: the steady torque that would heat the motor as much as the real varying torque. On: 05 Leg loads and stairs
• Robot description: Files that tell software what a robot is made of: links, joints, masses, inertias, shapes and limits. On: 13 The robot description: one source of truth
• Rod end: A ball joint at the end of a rod that lets it swivel in all directions while carrying push and pull forces. On: 06 The parallel-linkage ankle
• Root path: The trajectory of the hips over the ground (position and heading per frame), used as a constraint so a generated motion moves at a chosen speed and direction. On: 17 Motion tracking: Kimodo, retargeting and BeyondMimic
• ros2_control: Part of ROS 2 (a popular robot software framework) for talking to hardware in a tight, regular loop: it reads sensors, runs controllers and writes motor commands each cycle. On: 03 System architecture: from battery to policy

S

• Safety factor: The material’s yield strength divided by the peak stress in the part. 1 means the part just starts to bend permanently. On: 08 Structure and code-CAD
• Sagittal plane: The side-view plane that divides the body into left and right; forward-back and up-down motion lie in it. On: 01 Degrees of freedom and kinematic chains
• Second moment of area: A number (in mm⁴) that captures how far a beam’s cross-section material sits from its centre line. Material far from the centre resists bending much more, which is why I-beams and tubes are stiff for their weight. On: 08 Structure and code-CAD
• Semi-implicit Euler: An integrator that updates the velocity first, then moves with the new velocity; cheap, and unlike plain Euler it does not steadily gain energy over long runs. On: 14 Simulation: MuJoCo, Isaac Lab, Newton and PhysX
• Serial bus servo: A hobby-class servo with an ID that shares a daisy-chained serial line with others and reports its position and load. On: 19 Hands
• Serial mechanism: A chain of joints each driven by its own actuator, one after another, like a typical robot arm. On: 06 The parallel-linkage ankle
• Settling time: Time after a step until the joint stays within a band (here 2% of the step) of its final value. On: 15 PD control, gains and armature
• Sim-to-real gap: The difference between how the simulated robot and the real robot respond to the same commands, which can make a policy that works in simulation fail on hardware. On: 03 System architecture: from battery to policy, 14 Simulation: MuJoCo, Isaac Lab, Newton and PhysX, 20 Sim-to-real and system identification, 22 Build phases and roadmap
• Sim-to-sim: Checking a policy trained in one simulator in a second, independent one (Marvin Bot: trained on Newton, checked in MuJoCo), as a cheap proxy for the real world. On: 14 Simulation: MuJoCo, Isaac Lab, Newton and PhysX
• Sim-to-sim check: Running a policy trained in one simulator in a different one (Marvin Bot: Newton to MuJoCo) to catch dependence on one engine’s quirks. On: 20 Sim-to-real and system identification
• Single source of truth: Keeping each fact in one place and generating every other copy from it, so copies cannot disagree. On: 13 The robot description: one source of truth
• Singularity: A pose where joint axes line up so the mechanism loses a direction of motion; near it, small motions in that direction need very large joint speeds. Also called gimbal lock for three-axis joints. On: 01 Degrees of freedom and kinematic chains, 06 The parallel-linkage ankle
• SocketCAN: The part of Linux that makes a CAN adapter look like a network connection, so programs can send and receive CAN messages with ordinary networking tools. On: 10 CAN buses and bandwidth
• Soft contact: How MuJoCo handles contact: instead of perfectly rigid surfaces, it lets bodies overlap very slightly and push back like a stiff, damped cushion, which keeps the simulation stable. Settings called solref and solimp tune the cushion. On: 14 Simulation: MuJoCo, Isaac Lab, Newton and PhysX
• Spares: Extra units bought beyond what the robot uses, so that one failure does not stop the project for weeks. Marvin Bot keeps one spare RS04, RS03 and RS10P. On: 21 Bill of materials and costs
• Square-cube law: When a body is scaled up by s, its areas grow as s squared and its volume and mass as s cubed, so weight outgrows strength. Joint torques grow about as s to the fourth. On: 07 Mass budget and proportions
• Staged buying: Buying parts one build phase at a time, so a failed test costs only that phase’s parts. On: 21 Bill of materials and costs
• Stall torque: The torque a motor or servo produces when held still at full power; it can only be sustained for a moment before overheating. On: 19 Hands
• Stance (default pose): The standing pose the walking policies act around and start from: for Marvin Bot hip pitch 6°, knee 15°, ankle pitch 9°, with the pelvis 0.694 m above the floor. On: 01 Degrees of freedom and kinematic chains, 16 Reinforcement learning for walking
• State of charge: How full a battery is, from 0% to 100%. On: 09 Power and battery
• Steady-state error: The offset left after settling, e.g. gravity torque divided by kp for a PD loop. On: 15 PD control, gains and armature
• Stiffness (kp): The “P” setting of a PD controller (in N·m/rad): how hard a joint pulls back towards its target for each radian it is off, like the strength of a spring. On: 15 PD control, gains and armature
• Stop category: IEC 60204-1’s names for how a machine stops: 0 cut power at once, 1 controlled stop then cut power, 2 controlled stop with power kept. On: 12 Safety
• Stress concentration: A local rise in stress at a hole, notch or sharp corner, often 2-3 times the nominal value, expressed as a factor Kt. On: 08 Structure and code-CAD
• Support polygon: The smallest convex outline containing every point where the feet touch the ground. On: 02 Balance: centre of mass, support polygon, ZMP and capture point
• System 1 / System 2: GR00T’s split into two parts: a slower part that looks at images and reads the instruction to understand the scene (System 2, “thinking”), and a faster part that turns that understanding into motion (System 1, “acting”). On: 18 Whole-body control and VLA models
• System identification: Measuring a real system’s response to known inputs and fitting the parameters of a model so it reproduces that response. On: 20 Sim-to-real and system identification

T

• Teacher-student training: Train a “teacher” policy in simulation that is allowed to see perfect information (exact terrain, true speeds), then train a “student” that copies the teacher’s actions while seeing only what real sensors provide. On: 20 Sim-to-real and system identification
• Technology readiness level: NASA’s 1-9 scale of how proven a technology is, from basic principles up to a system proven in real use. On: 22 Build phases and roadmap
• Teleoperation: A person controlling the robot remotely (VR headset, exoskeleton, joystick), often to record demonstrations. On: 18 Whole-body control and VLA models
• Tendon drive: Moving finger joints by cables pulled by motors placed elsewhere, usually in the forearm or palm. On: 19 Hands
• TensorRT: NVIDIA’s inference optimiser and runtime that compiles a trained network for faster execution on a GPU or Jetson. On: 18 Whole-body control and VLA models
• Termination: A 120 ohm resistor at each end of a CAN bus that stops signal reflections. On: 10 CAN buses and bandwidth
• Thermal runaway: A lithium-ion cell heating itself uncontrollably, venting flammable gas and possibly igniting. On: 12 Safety
• Thermal time constant: How long a body takes to approach its final temperature (about 63% of the way after one time constant); windings take seconds, housings minutes. On: 04 Actuators: quasi-direct drive, torque, speed and heat
• Timestep: The fixed time between simulation updates; 2 ms in Marvin Bot’s MuJoCo checks, 5 ms in Isaac Lab training. On: 14 Simulation: MuJoCo, Isaac Lab, Newton and PhysX
• TOPS: Trillions of operations per second: the headline figure for an AI chip’s speed. It is usually quoted for low-precision maths and with a trick (sparsity) that doubles it, so real-world throughput is often half or less. On: 11 Sensors and compute
• Torque: The turning effect of a force: force times the perpendicular distance from the axis, in N·m. On: 05 Leg loads and stairs, 07 Mass budget and proportions
• Torque constant: Torque produced per ampere of motor current (N·m/A). RobStride quotes it at the gearbox output, e.g. 2.1 N·m/A for the RS04. On: 04 Actuators: quasi-direct drive, torque, speed and heat, 20 Sim-to-real and system identification
• Torque-speed line: The motor’s available torque falling from peak at standstill to zero at no-load speed; Isaac Lab’s DC motor model uses a straight line. On: 15 PD control, gains and armature
• Transmission angle: The angle between a crank and its rod; its sine (1 when square, 0 when in line) is the fraction of crank torque that becomes useful rod force. On: 06 The parallel-linkage ankle
• Tread: The horizontal depth of one stair step. On: 05 Leg loads and stairs

U

• Underactuation: Having fewer motors than joints; linkages, tendons or springs couple joints so one motor curls several and fingers wrap around objects. On: 19 Hands
• URDF: Unified Robot Description Format, the XML robot format of ROS; holds links, joints, inertia, visual and collision shapes and joint limits. On: 13 The robot description: one source of truth
• USD: Universal Scene Description (OpenUSD): the 3-D scene file format that NVIDIA Isaac Sim and Isaac Lab load. Extra physics information added to it describes the robot’s bodies and joints. On: 13 The robot description: one source of truth

V

• Value function: A helper network (the critic) that guesses how much total reward the robot will collect from a given situation. It is used only during training, to judge whether actions were better or worse than expected; it never runs on the robot. On: 16 Reinforcement learning for walking
• Viscous friction: Resistance proportional to speed, in N·m·s/rad. On: 20 Sim-to-real and system identification
• Vision-language-action model: A large neural network that takes camera images and a text instruction and outputs short chunks of robot motion; GR00T N1.7 is the one planned for Marvin Bot. On: 03 System architecture: from battery to policy
• Visual geometry: Shapes or meshes used only for drawing the robot; they have no effect on the physics. On: 13 The robot description: one source of truth
• VLA (vision-language-action model): A neural network that takes camera images, a text instruction and robot state and outputs robot actions. On: 18 Whole-body control and VLA models

W

• Watchdog: A circuit that trips if it stops receiving regular heartbeat signals from the controller; here a relay that opens the contactor. On: 12 Safety
• Whole-body control: Deciding all joint commands together so several goals (balance, hands, head, limits) are met at once, in order of priority. On: 18 Whole-body control and VLA models
• Workspace: The set of positions (and orientations) a link can reach within the joint ranges. On: 01 Degrees of freedom and kinematic chains

Y

• Yield strength: The stress at which a material starts to deform permanently. On: 08 Structure and code-CAD

Z

• Zero-moment point: The point on the ground where the combined push of gravity and the body’s motion acts, with no tendency to tip about it. While it stays inside the area under the feet the robot does not topple; on flat ground it is the same as the centre of pressure. On: 02 Balance: centre of mass, support polygon, ZMP and capture point
• Zero-order hold: A digital controller keeps its last output constant until the next update. On: 15 PD control, gains and armature