Glossary

This glossary collects technical terms used throughout the textbook, grouped by week.

Overview Terms

• optimal control problem: Finding control inputs that minimize a cost subject to system dynamics.
• dynamics: The rules/equations describing how state evolves over time under actions and disturbances.
• state: The minimal information needed to describe the system at a given time.
• control: Commands sent to the robot to influence future state.
• actuators: Physical components that convert control commands into motion/force.
• action: A decision variable selected by a policy (often at a higher abstraction than low-level control).
• reward: Scalar feedback used to measure desirability of outcomes in reinforcement learning.
• return: Cumulative (often discounted) sum of rewards over time.
• reinforcement learning: Learning a policy through interaction to maximize expected return.
• decision making under uncertainty: Choosing actions when state, outcomes, or models are uncertain.
• planning: Computing a sequence of actions/subgoals to achieve an objective.
• subgoal: Intermediate target used to decompose a larger task.
• setpoint: Desired target value/state for a controller to regulate to.
• post condition: Condition that should hold after executing a plan or action.
• perception: Inferring useful world/robot information from sensor data.
• sensor: Device that measures aspects of the robot or environment.
• uncertain: Not exactly known; represented with error bounds or probability.
• partial observability: Situation where the full true state cannot be directly observed.
• probabilistic belief: Distribution representing uncertainty about hidden state.
• inference (inferring): Estimating hidden variables from observations and models.
• sense-plan-act loop: Control architecture cycling through sensing, planning, and acting.
• perception-planning-control loop: Iterative architecture coupling state inference, decision making, and control execution.
• policy: Mapping from state/observation (or belief) to action/control.
• robust policy: Policy that maintains performance under uncertainty or model mismatch.
• adaptive policy: Policy that updates behavior as conditions/data change.
• autonomous system: System that perceives, decides, and acts with limited external intervention.
• agent: Entity that takes actions to optimize an objective.
• agency: Capacity of an entity to make decisions and produce actions.
• multi-agent system: Environment with multiple interacting decision-making agents.

Week 1: State Representations

• state: The minimal set of variables needed to describe a system at a given time.
• reference frame: A coordinate system used to express positions and orientations.
• world frame: A fixed global frame used as a common reference.
• robot frame: A frame attached to the robot body.
• orientation: The rotational part of pose.
• pose: Position plus orientation.
• rotation matrix: An orthonormal matrix that represents rotation between frames.
• elementary rotations: Axis-specific rotations (for example around x, y, z) used to build 3D orientation.
• non-commutative rotations: The property that rotation order changes the result.
• translation: Linear displacement by a vector.
• rigid body motion: Combined rotation and translation of a body with fixed shape.
• homogeneous transform: A matrix form that combines rotation and translation in one operation.
• transform chaining: Multiplying transforms to move through multiple frames.
• TF tree: A graph of coordinate-frame transforms used to represent robot geometry.
• metric map: A map with geometric distances and coordinates.
• cost map: A map layer assigning traversal or risk costs to space.
• point cloud: A set of 3D points (optionally with normals/features) representing geometry.
• signed distance field: A field storing distance to the nearest surface with inside/outside sign.
• occupancy grid: A discretized map where each cell stores occupancy probability.
• uncertainty: Lack of exact knowledge about state, measurements, or world structure.
• octree: A hierarchical tree structure for efficient 3D spatial partitioning.
• kd-tree: A data structure for partitioning and querying points in k-dimensional space.
• topological map: A graph-based map emphasizing connectivity over metric detail.
• scene graph: A structured graph representing objects/places and relations.
• local planning: Planning over a short horizon with local context.
• global planning: Planning over larger regions or full-map context.
• world model: A state representation that includes dynamics of the environment.

Week 2: Modelling

• configuration: The minimal variables that fully specify robot state in its configuration space.
• degrees of freedom (DoF): Number of independent coordinates needed to specify configuration.
• configuration space (C-space): The space of possible robot configurations.
• task space: The physical space where task-relevant behavior is expressed.
• manifold: A lower-dimensional structure embedded in a higher-dimensional space.
• fully actuated: System where all DoFs are directly controllable.
• underactuated: System where fewer independent controls exist than DoFs.
• motion model: A function describing state evolution under actions/inputs.
• discrete-time dynamics: State evolution represented at sampled time steps.
• continuous-time dynamics: State evolution represented by differential equations.
• disturbance: Unmodeled effect acting on system dynamics.
• bicycle model: Kinematic abstraction for car-like steering dynamics.
• differential drive: Two-wheel independent-drive mobile robot model.
• instantaneous center of rotation (ICR): The instantaneous pivot point of planar turning motion.
• unicycle model: Nonlinear planar mobile robot model with linear/angular velocity inputs.
• state-space form: Vector-matrix representation of dynamics.
• linearization: Local first-order approximation of a nonlinear model around an operating point.
• Jacobian: Matrix of first-order partial derivatives of a vector function.
• observation model: Mapping from latent state to sensor measurement.
• latent state: Underlying state variable not directly observed.
• measurement noise: Random sensor error in observations.
• latency: Delay between physical event and measurement availability.
• bias: Systematic measurement offset from true value.
• drift: Slow time-varying bias or accumulated error.
• time synchronization: Aligning sensor clocks/timestamps for consistent fusion.
• intrinsics: Internal sensor parameters (for example camera focal length, distortion).
• extrinsics: Rigid transform relating one sensor/body frame to another.
• calibration: Estimation of sensor/model parameters from data.
• outlier: Measurement inconsistent with assumed noise model.
• gating: Rejecting measurements based on statistical consistency thresholds.
• RANSAC: Robust estimator that fits models using random consensus subsets.
• Huber loss: Robust loss that is quadratic near zero and linear for large residuals.
• observability: Ability to infer hidden state from available measurements.

Week 3: Control

• open-loop control: Applying precomputed actions without feedback correction.
• feedback control: Using measured state/output to correct actions online.
• regulation: Driving state to a fixed target/setpoint.
• trajectory tracking: Following a time-varying reference trajectory.
• PID control: Proportional-Integral-Derivative feedback controller.
• proportional term (P): Control contribution proportional to current error.
• integral term (I): Control contribution from accumulated past error.
• derivative term (D): Control contribution from error rate of change.
• integral windup: Integrator accumulation causing excessive control after saturation/large transients.
• setpoint: Desired target state/value.
• saturation: Clamping control output to actuator limits.
• Linear Quadratic Regulator (LQR): Optimal linear-state feedback under quadratic cost.
• state error: Difference between current and desired/equilibrium state.
• Riccati equation: Matrix equation solved to compute optimal LQR gains.
• finite-horizon control: Optimization over a fixed time horizon.
• infinite-horizon control: Optimization over an unbounded horizon.
• iterative LQR (iLQR): Nonlinear trajectory optimization via repeated local LQR approximations.
• forward pass: Simulation rollout step in iterative trajectory optimization.
• backward pass: Dynamic-programming gain/correction computation in iLQR.
• feedforward term: Open-loop correction component computed by optimization.
• feedback gain: Closed-loop gain multiplying state deviation.
• Model Predictive Control (MPC): Receding-horizon optimization executed repeatedly online.
• receding horizon: Apply first optimized action, then re-solve at next step.
• state constraints: Allowed region limits on states.
• control constraints: Allowed region limits on inputs.
• quadratic program (QP): Optimization with quadratic objective and linear constraints.

Week 4: State Estimation

• ground truth: True but typically inaccessible state.
• belief: Probability distribution representing state uncertainty.
• likelihood: Probability of an observation given a hypothesized state.
• prior: Belief before incorporating current measurement.
• posterior: Updated belief after applying Bayes rule.
• Bayes rule: Formula combining prior and likelihood to produce posterior.
• marginal likelihood (evidence): Normalizing probability of the observation.
• recursive Bayes filter: Predict-update recursion for sequential state estimation.
• prediction step: Propagate belief through transition/motion model.
• update step: Correct predicted belief using measurement likelihood.
• transition model: Probabilistic dynamics model for state evolution.
• histogram filter: Discrete Bayes filter over finite state bins.
• normalization constant: Factor that ensures probabilities sum/integrate to one.
• Kalman filter (KF): Optimal linear-Gaussian recursive estimator.
• mean: Expected value of a probability distribution.
• covariance: Second-moment uncertainty and correlation structure.
• innovation/residual: Difference between observed and predicted measurement.
• innovation covariance: Uncertainty of residual used for weighting measurements.
• Kalman gain: Optimal weighting between prediction and measurement.
• Extended Kalman Filter (EKF): Nonlinear estimator using local linearization.
• particle filter: Sample-based Bayesian estimator with weighted particles.
• resampling: Re-drawing particles to avoid weight degeneracy.
• sequential Monte Carlo: Family of particle-based sequential inference methods.
• sensor fusion: Combining multiple sensor streams in one estimation framework.

Week 5: Navigation and Mapping

• SLAM: Simultaneous localization and mapping.
• joint state: Combined state vector containing robot and map variables.
• loop closure: Reduction of accumulated uncertainty by re-observing known places/features.
• landmark: Distinct map feature used for localization/mapping.
• data association: Matching current observations to existing map entities.
• Mahalanobis distance: Covariance-aware distance used for statistical matching.
• log-odds: Logit representation of occupancy probability used for additive updates.
• ray casting: Tracing sensor rays through map cells for occupancy updates.
• Rao-Blackwellization: Factorization using conditional independence to reduce inference complexity.
• FastSLAM: Particle-based SLAM with per-particle landmark estimators.
• global planner: Planner over full map for long-horizon path generation.
• local planner: Short-horizon reactive planner for immediate feasible motion.
• Bug algorithms: Reactive obstacle-boundary-following navigation family.
• M-line: Straight line from start to goal used by Bug2.
• A*: Heuristic graph-search algorithm minimizing path cost.
• heuristic admissibility: Heuristic property that guarantees no overestimation of true cost.
• D*: Incremental replanning algorithm for changing/partially known maps.
• probabilistic roadmap (PRM): Sampling-based graph planner for high-dimensional spaces.
• Rapidly-exploring Random Tree (RRT): Incremental sampling-based tree planner.
• RRT*: Asymptotically optimal RRT variant with rewiring.
• Dynamic Window Approach (DWA): Local velocity-space planner under dynamic constraints.
• admissible velocity set: Velocities that allow safe stopping before collision.
• trajectory rollout: Simulating candidate controls over short horizon and scoring outcomes.
• tentacle planner: Local planning using precomputed candidate arcs.
• navigation stack: Integrated pipeline of estimation, mapping, planning, and control.

Week 6: Articulated Robots and Kinematics

• articulated chain: Rigid links connected by joints.
• link: Rigid body element in a robot mechanism.
• joint: Mechanical connection allowing constrained relative motion.
• revolute joint: Joint with rotational DoF.
• prismatic joint: Joint with translational DoF.
• end-effector: Terminal link/tool that interacts with task environment.
• Denavit-Hartenberg (DH) convention: Standard parameterization for serial-chain kinematics.
• DH parameters: Geometric parameters (a, alpha, d, theta) defining adjacent-link transforms.
• URDF: Unified Robot Description Format for robot structure, geometry, and properties.
• visual geometry: Mesh/shape used for rendering robot appearance.
• collision geometry: Simplified shape used for collision checks.
• inertial parameters: Mass properties including COM and inertia tensor.
• forward kinematics (FK): Compute end-effector pose from joint configuration.
• workspace: Set of poses reachable by some configuration.
• inverse kinematics (IK): Compute joint configuration that achieves desired pose.
• kinematic redundancy: More DoFs than needed for primary task constraints.
• analytical IK: Closed-form symbolic IK solution for specific robot geometries.
• numerical IK: Iterative optimization/linearization-based IK.
• resolved-rate control: IK strategy using differential kinematics and Jacobian.
• Jacobian pseudoinverse: Least-squares inverse mapping from task velocity to joint velocity.
• singularity: Configuration where Jacobian loses rank and motion authority degrades.
• joint limits: Physical bounds on joint positions/velocities.

Week 7: Dynamics

• manipulator equations: Standard joint-space dynamics equation for robot arms.
• mass matrix: Configuration-dependent inertia mapping from acceleration to torque.
• Coriolis/centripetal term: Velocity-dependent dynamic coupling term.
• gravity vector: Joint torques required to counter gravity.
• joint torque: Actuation torque applied at joints.
• forward dynamics: Compute accelerations from current state and applied torques.
• inverse dynamics: Compute torques required for desired motion.
• recursive Newton-Euler algorithm: O(n) method for inverse dynamics of serial chains.
• computed torque control: Model-based controller canceling dynamics and adding feedback.
• dynamic decoupling: Transforming coupled nonlinear dynamics into simpler independent forms.
• system identification: Estimating model parameters from motion/force data.
• regressor form: Dynamics expressed linearly in unknown parameters.
• persistently exciting trajectory: Motion rich enough to identify parameters reliably.
• robot calibration: Estimation of geometric model parameters for accuracy.
• contact dynamics: Dynamics when bodies interact through constraints and forces.
• unilateral constraint: Contact can push but not pull.
• complementarity condition: Mathematical condition coupling contact gap and normal force.
• Coulomb friction model: Friction cone model limiting tangential force by normal force.
• Linear Complementarity Problem (LCP): Optimization/feasibility form common in rigid contact simulation.
• sim-to-real gap: Performance drop when transferring from simulation to physical robot.
• domain randomization: Randomizing simulation parameters during training for robustness.

Week 8: Perception and Learning

• robot learning: Learning perception, dynamics, or policy components from data.
• supervised learning: Learning from labeled input-output examples.
• dataset: Collected examples used for model fitting and validation.
• loss function: Objective minimized during model training.
• classification: Predicting discrete labels from inputs.
• regression: Predicting continuous outputs from inputs.
• segmentation: Predicting per-pixel or per-point labels.
• feature representation: Learned or engineered input abstraction for downstream tasks.
• latent variable: Hidden variable inferred from observations.
• dynamics learning: Learning transition models from state-action-next-state data.
• residual dynamics model: Learned correction added to known physics model.
• probabilistic dynamics model: Transition model that outputs uncertainty/distribution.
• imitation learning: Learning policy behavior from demonstrations.
• behavior cloning: Supervised imitation mapping observations directly to expert actions.
• covariate shift: Distribution mismatch between training states and deployment states.
• DAgger: Dataset Aggregation method reducing imitation-learning distribution shift.
• action representation: Chosen control parameterization predicted by a learned policy.
• stochastic policy: Policy outputting a distribution over actions.
• mixture density network (MDN): Model outputting mixture distributions for multimodal targets.
• diffusion policy: Policy that samples actions through iterative denoising.
• motion primitive: Structured reusable movement pattern.
• Dynamic Movement Primitive (DMP): Stable attractor-based movement primitive with learned forcing term.
• probabilistic movement primitive: Distributional extension of movement primitives.
• Decision Transformer: Sequence model treating control as conditional sequence prediction.
• vision-language-action (VLA) model: Policy conditioned on visual observations and language commands.
• inductive bias: Structural assumptions that improve generalization/data efficiency.
• closed-loop evaluation: Testing learned models in feedback execution, not only offline metrics.

Week 9: Reinforcement Learning

• reinforcement learning (RL): Learning policies from interaction to maximize long-term reward.
• agent: Decision-maker that acts in an environment.
• environment: External process generating transitions and rewards from actions.
• Markov Decision Process (MDP): Formal model of fully observable sequential decision making.
• state space: Set of possible states in an MDP.
• action space: Set of allowable actions.
• transition probability: Probability of next state conditioned on current state/action.
• reward function: Scalar signal evaluating transitions/actions.
• discount factor: Weighting factor for future rewards.
• Markov property: Future depends only on present state and action, not full history.
• return: Discounted cumulative future reward.
• Partially Observable MDP (POMDP): Decision model with hidden true state and partial observations.
• belief state: Probability distribution over hidden states.
• value function: Expected return from a state under a policy.
• action-value function (Q-function): Expected return from taking action in state then following policy.
• Bellman equation: Recursive relation defining value functions.
• Bellman optimality equation: Recursive condition for optimal value.
• temporal-difference (TD) learning: Value learning via bootstrapped one-step targets.
• TD error: Difference between predicted value and bootstrapped target.
• dynamic programming: Exact MDP solution methods with known model.
• value iteration: Bellman-optimality fixed-point iteration over values.
• policy iteration: Alternate policy evaluation and policy improvement.
• Q-learning: Off-policy TD method for learning optimal Q-function without model.
• off-policy learning: Learning one policy while following another behavior policy.
• epsilon-greedy exploration: Random-action exploration mixed with greedy exploitation.
• Deep Q-Network (DQN): Neural-network approximation of Q-values.
• experience replay: Random minibatch training from stored transitions.
• target network: Delayed Q-network used to stabilize training targets.
• actor-critic: RL framework with separate policy (actor) and value estimator (critic).
• advantage: Relative action quality compared to baseline state value.
• Proximal Policy Optimization (PPO): Policy-gradient method with clipped updates for stability.
• sim-to-real transfer: Deploying policies trained in simulation onto hardware.
• reward engineering: Designing reward functions that produce intended behavior.
• credit assignment: Determining which past actions caused delayed outcomes.

Week 10: Human-Robot Interaction (HRI)

• human-robot interaction (HRI): Design and analysis of robot behavior with/around people.
• human factors: Human capabilities/limitations relevant to system design.
• proxemics: Social use of interpersonal distance.
• BDI model: Belief-Desire-Intention cognitive architecture for goal-directed agents.
• intent inference: Estimating a human's likely goal/purpose from observations.
• human motion prediction: Forecasting future human trajectories under uncertainty.
• Social Force Model: Crowd/pedestrian model using attractive/repulsive interaction forces.
• inverse reinforcement learning (IRL): Inferring reward/objective from demonstrated behavior.
• safety-rated monitored stop: Safety mode where robot halts when humans enter defined area.
• speed and separation monitoring: Adjusting robot speed based on human distance.
• power and force limiting: Collaborative safety mode limiting collision forces.
• compliance control: Control that yields under interaction forces instead of rigidly resisting.
• impedance control: Force-motion behavior shaped as virtual mass-spring-damper.
• chance constraint: Probabilistic safety constraint (for example collision probability bound).
• shared control: Human and robot jointly influence action execution.
• predictability: Ease with which humans can anticipate robot behavior.
• legibility: Robot behavior designed to communicate intent clearly.
• transparency: Clarity of robot internal state, intention, and decision rationale to users.
• trust calibration: Aligning human trust with true robot capability.
• technology acceptance: User willingness to adopt/use a system.
• Uncanny Valley: Drop in comfort/acceptance for near-human-but-not-quite agents.
• multi-agent system: System of interacting decision-making agents.
• task allocation: Assignment of subtasks across humans and robots.
• coordination: Temporal and behavioral alignment across team members.
• situation awareness: Understanding of environment, task state, and team status.
• handover: Coordinated transfer of object/control between human and robot.
• affective computing: Modeling or recognizing emotion/affective state in interaction.
• personalization: Adapting robot behavior to individual user preferences/capabilities.
• Wizard of Oz (WoZ) study: HRI study where hidden human operator simulates autonomy.
• ecological validity: Extent to which study findings generalize to real-world settings.
• NASA-TLX: Standard questionnaire for perceived workload measurement.