Glossary

This glossary defines key terms used throughout Physical AI. Terms are organized alphabetically.

A

Action Chunking: A policy execution technique where a model predicts an open-loop sequence of future actions over a multi-step horizon rather than a single action per step. Amortizes the latency of slow neural inference across multiple control intervals and dampens high-frequency compounding covariate shift by generating kinematically coordinated trajectories. In closed-loop deployment, overlapping chunks are executed in a receding-horizon fashion or blended via temporal ensembling to maintain responsiveness to dynamic environmental disturbances. (The Cognitive Brain, Trajectory Planning, Policy Training)

B

Behavioral Cloning (BC): The foundational imitation learning paradigm that trains a policy via supervised learning to directly map sensory observations to expert control actions, treating physical sequential decision-making as a direct regression or classification task. While computationally efficient on static demonstration archives, standard behavioral cloning ignores the closed-loop consequences of its own control outputs. As a result, small single-step inference errors induce compounding covariate shift, driving the physical robot away from the demonstration manifold into unrecoverable failure states. (Policy Training, Physical Data)
Belief State Invalidation Horizon (\(t_{\text{exp}}\)): The elapsed temporal duration beyond which an unobserved spatial or kinematic state estimate can no longer be certified to lie within a task’s safe tolerance envelope. Evaluated from environmental disturbance dynamics, unmodeled obstacle velocities, and estimator drift, it establishes the expiration shelf life of perceptual memory. Downstream planners that treat spatial beliefs older than this horizon as valid free space risk commanding collisions with dynamic obstacles. (Spatial Memory, Continuous Covariance Propagation and Observability)
Bumpless Transfer: A dynamic control handoff protocol that matches state, velocity, and effort references across control authorities at the instant of switching, eliminating transient step changes in commanded actuator torques. When supervisory control transitions abruptly between an autonomous neural policy and a human operator, raw authority switching induces mechanical shocks, structural oscillations, and loss of tire traction. Bumpless transfer employs dynamic transition ramps and feedback state tracking to smoothly transfer control energy within the actuator’s physical bandwidth. (Supervisory Intervention, Smooth Trajectories and Bumpless Transfer)

C

Causal Boundary: The physical, energetic, and architectural threshold in an embodied system where reversible digital computations terminate and irreversible physical work is performed on the external environment (\(dW = \mathbf{F} \cdot d\mathbf{x}\)). Beyond this boundary, computational errors cannot be rolled back by software exception handlers or abort routines, requiring that safety invariants be verified before actuator register writes occur. The permission path sits above this boundary, at the proposal boundary of The Machine in Five Levels. (The Causal Boundary, The Physical Body)
Causal Confusion (Causal Misidentification): An imitation learning failure mode where a policy conditions on spurious, non-causal features that correlate with expert actions in training demonstrations rather than the true physical affordances of the task. High-capacity neural models exploit incidental visual cues (such as interface indicators, dashboard lights, or past actions) that disappear or change under physical deployment. Deployed closed-loop policies fail catastrophically when test-time interventions break these non-causal correlations. (Policy Training)
Clearance Inset Margin: A deterministic geometric safety buffer subtracted from perceived free-space boundaries to guarantee collision avoidance in the presence of spatial measurement uncertainty. The inset is formally computed from the vector sum of sensor calibration error, depth covariance, transform chain kinematic tolerances, and dynamic tracking error envelopes. Enforcing a clearance inset prevents a planner from routing trajectories along the boundary of raw sensor detections, ensuring physical clearance under worst-case perception drift. (Sensor Perception, Trajectory Planning)
Closed-Loop Evaluation: The metrological paradigm in physical AI that measures a policy’s performance, safety, and stability through interactive rollouts on physical hardware or high-fidelity simulation, rather than through static loss metrics on a held-out dataset. Because an autonomous policy continuously perturbs its own operating environment, meaningful evaluation requires logging closed-loop metrics such as task completion rates, dynamic safety margin violations, and mean operational time between interventions. Evaluating policies purely via open-loop validation error obscures catastrophic compounding covariate shift. (Closed-Loop Evaluation, Policy Training)
Compounding Rollout Error: The theoretical and empirical accumulation of trajectory deviations in open-loop imitation learning, where a per-step policy error rate of \(\epsilon\) induces a total trajectory error that compounds quadratically (\(\mathcal{O}(\epsilon T^2)\)) over execution horizon \(T\). Because the agent encounters unseen states as soon as it deviates from the demonstration path, its conditional error probability spikes rather than remaining constant. In physical robotics, this quadratic growth transforms minor tracking drift into rapid collisions with obstacles or mechanical joint-limit saturation. (Policy Training, Imitation Learning and Compounding Covariate Shift)
Control Barrier Function (CBF): A continuously differentiable scalar certificate \(h(\mathbf{x}) \ge 0\) used to define and guarantee the forward invariance of an admissible safe set in the state space of a dynamical system. By enforcing the boundary rate-of-approach condition \(\dot{h}(\mathbf{x}, \mathbf{u}) \ge -\alpha(h(\mathbf{x}))\), a CBF translates complex state constraints into instantaneous affine inequality constraints on control inputs. This allows candidate proposals from unverified neural policies to be projected onto safe control inputs in real time via convex quadratic programming. (Safety Enforcement, Safety and Control Barrier Functions)
Cross-Embodiment Transfer: A learning paradigm in which a single policy or foundation model is trained on sensorimotor data gathered across heterogeneous robot morphologies, kinematic chains, and actuator dynamics to generalize to novel target embodiments. Achieving cross-embodiment generalization requires normalizing disparate observation spaces into unified coordinate frames and mapping heterogeneous action spaces through embodiment-agnostic representations. Its primary challenge is preventing negative transfer induced by unmodeled dynamic discrepancies, link inertia variations, and differing joint velocity boundaries. (Physical Data, Policy Training)

D

Diffusion Policy: A generative policy architecture that models continuous, multi-modal robot action trajectories by inverting a parameterized stochastic diffusion process conditioned on sensory observations. By learning the score function of the demonstration distribution, diffusion policies generate smooth, multi-step action chunks from Gaussian noise without collapsing into mode averaging or requiring explicit discretization of continuous control spaces. Their expressive density modeling makes them exceptionally effective for complex contact-rich manipulation tasks. (Policy Training, Continuous trajectory diffusion and flow matching)
Domain Randomization (DR): A sim-to-real transfer methodology that systematically perturbs physical parameters (mass, friction, damping, motor torque limits), sensor characteristics (noise covariance, latency, camera pose), and visual textures across simulated training episodes. By exposing the policy to a broad distribution of simulated environments, domain randomization encourages the neural network to learn invariant representations and robust feedback behaviors that treat the unmodeled physical world as another randomized instance within its training support. If randomized too narrowly, the real world falls outside the training hull; if randomized too broadly, policy performance degrades due to excessive conservatism. (Policy Training)
Dynamic Stopping Distance (\(d_{\text{stop}}\)): The clearance a moving embodied system needs to reach standstill from its current speed under credible braking and the full pre-brake delay: \(d_{\text{stop}} = v\,\tau_{\text{delay}} + \frac{v^2}{2 a_{\text{brake}}} + \delta_{\text{loc}} + \delta_{\text{margin}}\). It comprises the distance traveled during the delay from detection to retarding force, which grows linearly with that delay, the braking distance under a validated deceleration floor, which grows with the square of speed, and fixed allowances for localization error and protective clearance. The permission path admits a command only if this distance, inset by the tracking bound, fits within the clear distance to the boundary. (The Physical Body, Safety Enforcement)

E

Endogenous Covariate Shift: A distribution shift in sequential decision-making where an embodied agent’s own execution errors drive its physical body into states unrepresented in its training demonstrations, thereby presenting the policy with novel out-of-distribution observations. Because offline supervised training assumes independent and identically distributed data, policy errors under endogenous shift compound along the trajectory rather than averaging to zero. In physical systems, this shift leads to task failure or boundary violations unless mitigated by closed-loop corrective data or receding-horizon re-planning. (The Causal Boundary, Physical Data, Policy Training)
Endogenous Experience (The Action–Observation Loop): The closed-loop data generation process unique to embodied systems wherein an agent’s executed actions causally determine its subsequent physical states and sensory observations (\(s_{t+1} \sim P(s \mid s_t, a_t)\)). Unlike passive supervised learning on fixed, exogenous datasets, physical AI experience generation forms an irreversible dynamical feedback loop where minor control perturbations alter future data collection trajectories. Consequently, data generation and policy deployment are fundamentally inseparable in embodied agents. (The Causal Boundary, Physical Data)
Envelope Trichotomy: A three-valued epistemic classification of cyber-physical runtime health that partitions system status relative to its certified operating envelope into known-true (verified valid telemetry inside bounds), known-false (verified limit breach), and unknown (epistemic uncertainty from sensor corruption, frame drops, or estimator divergence). In embodied AI, both known-false and unknown strictly withhold tracking permission and trigger safe fallbacks, preventing unmeasured kinetic energy from accumulating during perceptual blind spots. This departs fundamentally from disembodied software, which often presumes execution safety in the absence of explicit error flags. (Deployment Release)
Exposure Wall: The limit on what failure-free operation alone can establish about a rare failure rate. Under a stationary Poisson failure model, bounding a catastrophic rate \(\lambda = 10^{-9}\) per hour at 95 percent confidence requires about \(-\ln(0.05)/\lambda \approx 3.0 \times 10^9\) failure-free hours of exposure from the same configuration, an argument Butler and Finelli made for flight software. A zero-event test cannot reach such a rate, so claims at that level rest on architectural evidence, fault injection, and runtime monitors rather than on accumulated operating hours alone. (Physical Trial Limits, Adversarial Verification, Zero-Failure Testing and the Exposure Wall)

F

Fallback Ladder: The ordered set of responses the permission path selects when a proposal cannot be admitted unchanged or its premises fail. Its rungs are PROJECT (minimal-intervention projection of the proposal onto the safe set), HOLD (a state-matched powered hold), STOP (a controlled, state-matched stop), and INHIBIT (drive torque inhibit with a separately rated brake). Each rung is selected before the admissible set becomes empty, and each is a validated response for the declared load and timing, not a generic abort. (The Fallback Ladder, Supervisory Intervention)
Flow Matching Policy: A continuous generative policy architecture that trains a neural vector field to transform a base probability distribution into an empirical action distribution along deterministic, straight-line optimal transport paths. Compared to score-based diffusion policies, flow matching models achieve comparable distribution expressiveness while requiring significantly fewer numerical integration steps to generate an action chunk. This inference speedup substantially reduces on-device generation latency, making real-time high-frequency closed-loop execution viable on embedded edge computing silicon. (Policy Training)
Forward Invariance: A mathematical property of a dynamical system and safe set \(\mathcal{C}\) asserting that any state trajectory initialized within \(\mathcal{C}\) remains strictly confined within \(\mathcal{C}\) for all future time under an admissible control policy. In physical AI, forward invariance serves as the foundational mathematical formulation of safety, guaranteeing that an embodied agent never breaches kinematic, dynamic, or thermal hazard boundaries during closed-loop operation. It is certified at runtime by verifying that control inputs satisfy subtangential boundary conditions at the perimeter of the safe set. (Safety Enforcement, Safety and Control Barrier Functions)

G

Grounded Affordance: The spatial, kinematic, and dynamic interaction manifold associated with an object or terrain region that defines where and how an embodied agent can initiate, sustain, and complete a physical manipulation or locomotion primitive. Unlike purely semantic object labels, a grounded affordance parameterizes contact surface normals, friction cone bounds, end-effector approach corridors, and kinematically feasible grasp poses relative to the robot’s physical embodiment. A semantic task request cannot be translated into physical action until its underlying affordance is rigorously grounded within the robot’s reachable workspace. (Grounded Intent, Sensor Perception)

H

Heterogeneous Silicon Partitioning: A hardware-enforced allocation strategy that physically isolates stochastic, non-deterministic machine learning workloads (running on GPUs or NPUs under general-purpose operating systems) from deterministic, safety-critical reflex loops (running on bare-metal or RTOS cores). By enforcing hardware memory protection boundaries and dedicated peripheral access, it eliminates common-mode OS panics and memory contention between neural pipelines and microsecond motor control loops. This ensures that the permission path retains guaranteed execution budgets and unshared bus access under all computational loads. (Silicon Placement, Heterogeneous SoC Mailboxes and Memory Barriers)

I

Inevitable Collision State (ICS): A physical system state \(\mathbf{x}(t)\) from which no admissible, physically executable control trajectory exists—within dynamic actuator and friction limits—that can prevent impact with environmental obstacles over an infinite time horizon. Unlike geometric obstacle clearance, an ICS explicitly accounts for plant momentum, friction limits, dynamic obstacle trajectories, and finite actuator braking authority. A physical AI agent enters an ICS long before geometric contact occurs if its velocity exceeds its stopping authority within available clearance. (The Physical Body, Trajectory Planning, Safety Enforcement)
Intent Grounding Refusal: A deterministic safety rejection mechanism whereby an intent translation engine or admittance gate formally refuses to execute a high-level command because the requested goal cannot be physically or safely grounded. Refusal is triggered when semantic targets lack reachable affordance manifolds, demand trajectories that violate kinodynamic feasibility, or conflict with verified safety invariants. Rather than silently attempting an under-actuated or hazardous motion, the system emits an explicit refusal diagnostic, signaling supervisory intervention or deliberative replanning. (Grounded Intent, Supervisory Intervention)
Intent Lease: A time-bounded and state-bounded authorization token under which a high-level cognitive goal or target affordance remains valid for physical execution by downstream planners. Because dynamic environments continuously evolve, an unfulfilled intent becomes physically obsolete or dangerous if delayed beyond a certified temporal lease or if the underlying scene evidence drifts beyond specified spatial tolerances. Expiration of the intent lease forces an immediate deliberative renegotiation or a safe transition to a standstill, preventing the robot from pursuing stale perceptual premises. (Grounded Intent, Trajectory Planning)
Intervention Masking: An empirical evaluation and safety pathology where repeated, proactive interventions by human supervisors systematically distort autonomy metrics by intercepting the policy before it encounters critical boundary failures. Because the supervisor intervenes during nascent drift, the policy’s recorded telemetry logs contain artificially elevated success rates and zero observed accidents, while hiding the true underlying failure rate and preventing the collection of recovery data. Unless explicitly accounted for via intervention-discounted metrics, intervention masking creates a false perception of policy readiness that leads to catastrophic failures upon unsupervised deployment. (Supervisory Intervention, Closed-Loop Evaluation)
Intervention Rate (Mean Time Between Interventions): The standardized reliability and safety metric in physical AI defined as the total number of supervisory human takeovers or emergency safety trips divided by the cumulative operating time or operational distance across physical rollouts. Unlike binary task success rates, MTBI provides an un-censored, statistically rigorous measure of an embodied agent’s autonomous survival capacity in open-world environments. Tracking MTBI exposes whether a high task completion rate is an illusion created by frequent human safety interventions that prevent the policy from encountering its failure boundaries. (Closed-Loop Evaluation, Supervisory Intervention)

K

Kinodynamic Feasibility: The mathematical property of a planned physical trajectory that simultaneously satisfies geometric configuration-space collision avoidance constraints, kinematic mechanism limits (joint positions and velocities), and dynamic plant boundaries (actuator torques, joint accelerations, reflected jerk, and contact friction cones). A purely kinematic trajectory that ignores actuator torque-speed curves or vehicle tire adhesion limits is physically unexecutable and will induce actuator saturation, phase lag, and control runaway. Trajectory synthesis in physical AI requires verifying kinodynamic feasibility across the entire motion horizon prior to execution. (Trajectory Planning, The Physical Body)

M

Metric 3D Lifting (Spatial Grounding): The geometric and learning process of projecting uncalibrated 2D image-space features, masks, or semantic tokens into metric, Euclidean three-dimensional representations anchored to the physical robot’s coordinate frame. Methods such as Ray Frustum Voxel Lifting, Truncated Signed Distance Function fusion, and 3D Gaussian Splatting convert scale-ambiguous pixel patterns into metric obstacle boundaries and free-space occupancy grids. Spatial grounding ensures that high-level neural semantic claims correspond to physically reachable, collision-checkable coordinate manifolds in the workspace. (Sensor Perception, Grounded Intent)
Minimal Risk Maneuver (MRM): An autonomous, deterministic fail-safe trajectory executed upon critical fault detection or loss of operational envelope to transition an embodied system from an active mission state into a stationary, low-energy safe state. Unlike an abrupt unpowered stop, which can induce dynamic tip-over, loss of steering control, or payload release, an MRM actively executes a controlled deceleration, lane clearance, or compliant ground settling. It represents the terminal behavioral guarantee of the permission path when nominal tracking authority is permanently revoked. (Safety Enforcement, Supervisory Intervention)
Minimal-Intervention Safety Shield: A real-time supervisory projection filter that mediates between an unverified learned policy and physical actuator drives by solving a quadratic program on every control cycle: \(\mathbf{u}^* = \arg\min_{\mathbf{u}} \|\mathbf{u} - \mathbf{u}_{\text{nom}}\|^2\) subject to barrier and actuator constraints. It passes nominal neural network proposals unaltered whenever they respect safe dynamics, intervening only at the boundary of the safe set to minimally redirect control commands along the tangent cone of forward invariance. This guarantees physical safety while preserving maximum task performance and behavioral expressiveness of the underlying model. (Safety Enforcement, Safety and Control Barrier Functions)
Mode Averaging (Multimodality Catastrophe): A catastrophic learning failure that occurs when a unimodal policy parameterized with a standard regression loss (such as mean squared error) is trained on multi-modal demonstration data exhibiting multiple valid pathways. When human demonstrators navigate around an obstacle by choosing either a left or right detour, a naive regression model averages the competing action modes, commanding an interpolation that steers the robot directly into the obstacle. Resolving mode averaging requires expressive generative decoders capable of modeling arbitrary multi-modal distributions over continuous action trajectories. (Policy Training, The Cognitive Brain)
Mode Confusion: A safety-critical cognitive misalignment between a human operator and an automated system regarding which entity possesses active control authority over specific physical degrees of freedom. Mode confusion arises when autonomous overrides engage or disengage silently without salient multisensory feedback, leading the human to believe they are steering while the autonomous policy is resisting, or vice versa. In safety-critical robotics, mode confusion induces destructive control conflicts, prolonged takeover reaction latencies, and accidental disengagement of critical safety shields. (Supervisory Intervention)
Monitored Premise Expiration: The automated runtime revocation of safety case validity triggered when empirical operational conditions (such as friction coefficients, ambient temperature, sensor packet loss, or battery internal impedance) violate the core design premises under which the system was qualified. By linking runtime telemetry monitors directly to safety case warrants, it prevents silent degradation of physical margins from exposing the machine to unshielded hazards. When a monitored premise expires, the system transitions autonomously into a restricted or refusal operational regime. (Deployment Release)

O

Observation Contract: A formal, deterministic specification establishing the precise spatial, temporal, and semantic guarantees that a perception subsystem commits to providing downstream planning and safety modules. The contract defines coordinate frame conventions, maximum permissible spatial information age, minimum depth resolution, and bounding covariance ellipsoids for detected obstacles. If dynamic occlusions, sensor dropouts, or computational bus contention cause perception output to violate these declared bounds, the contract triggers an immediate perceptual failure flag rather than transmitting stale or unverified spatial claims. (Sensor Perception, The Cognitive Brain)
Observational Indistinguishability: A fundamental perceptual limitation wherein two distinct physical world states—one nominal and one catastrophic—produce identical sensor telemetry up to the machine’s physical intervention deadline (\(t_{\text{deadline}} = t_{\text{harm}} - t_{\text{act}}\)). When sensor physics provides zero distinguishing energy before the time-to-harm expires, no software algorithm, neural capacity scaling, or anomaly detector can prevent failure. Mitigating observational indistinguishability demands conservative mechanical bounds, passive physical compliance, or scheduled non-destructive inspection rather than deeper neural networks. (The Epistemic Frontier)
Operational Design Domain (ODD): The object poses, surface conditions, payloads, speeds, and disturbances in which a machine is meant to operate, written down as the scenario ledger. A dataset supports the part of the ODD that its occupancy join covers; the rest is its known absence. The sets nest: the policy manifest claims a declared ODD within the ODD, evaluation samples a target ODD within the declared ODD, and verification extends the target ODD to the machine’s internal conditions as the operating envelope that a release verdict authorizes in whole, in part, or not at all. (Collection Policy Coverage, Policy Training, Closed-Loop Evaluation, Adversarial Verification, Deployment Release)

P

Physical Artificial Intelligence (Physical AI): An interdisciplinary branch of artificial intelligence concerned with computational systems whose sensory inputs, inferential deliberations, and learning algorithms are intrinsically coupled to physical bodies operating within continuous, non-stationary physical environments. Unlike disembodied AI systems that operate on static digital tokens, a physical AI system must continuously reconcile neural inference latency with real-world dynamical time scales, irreversible thermodynamic contacts, and hardware safety constraints. Its central objective is the synthesis of closed-loop policies that achieve generalizable task execution while strictly preventing catastrophic physical harm. (The Causal Boundary, The Cognitive Brain)
Proposal Boundary: The line between a learned model, which may only propose an action, and the permission path, an independent mechanism with bounded latency that the model cannot delay or corrupt and that alone may let the action reach an actuator. The permission path checks each proposal against fresh evidence, the stopping envelope, and a fallback feasible from the current state, and it refuses when any check fails. The proposal boundary separates the Brain from the Nervous System in the book’s five-level machine. (The Machine in Five Levels, The Nervous System, Safety Enforcement)

Q

Qualification Ladder (XIL): A structured cyber-physical verification progression that systematically tests embodied AI systems across evaluation substrates of increasing physical and temporal realism: Model-in-the-Loop (MIL), Software-in-the-Loop (SIL), Processor-in-the-Loop (PIL), Hardware-in-the-Loop (HIL), and In-Situ Physical Fault Injection. Because each stage closes or substitutes distinct physical and computational feedback loops, higher rungs provide authentic bus contention, electrical transients, and actuator dynamics but sacrifice the combinatorial test throughput of virtual rungs. Comprehensive assurance requires navigating the full ladder rather than relying on any single testing environment. (Adversarial Verification)

S

Safe Torque Off (STO): A certified hardware safety function (IEC 61800-5-2) that electronically disables pulse-width modulation gate signals to motor inverter power transistors, preventing the generation of electromagnetic rotational torque. Serving as the foundation for Category 0 emergency stops, STO guarantees that no software malfunction, memory corruption, or processor lockup can command motor actuation, leaving the plant to coast or rely on passive mechanical brakes. It serves as the ultimate non-software-dependent hardware fail-safe in the physical AI stack. (Safety Enforcement, The Physical Body)
Safety Case (Claim–Argument–Evidence): A structured, auditable assurance argument that demonstrates how an embodied AI system satisfies its explicit safety requirements within a declared operational design domain. It organizes assurance hierarchically by linking high-level claims to empirical and formal evidence (such as fault injection manifests, HIL qualification results, and barrier proofs) via transparent, falsifiable engineering arguments. In physical AI, the safety case explicitly binds software performance to physical hardware limits and operational envelopes. (Deployment Release)
Shared Autonomy: A collaborative control architecture in which a human supervisor and an autonomous policy concurrently exert control authority over an embodied system’s physical degrees of freedom. Authority allocation may be realized via continuous haptic virtual fixtures, dynamic compliance blending, or task-space privilege partitioning. The central challenge of shared autonomy is maintaining an intuitive, transparent authority distribution that prevents human-robot fighting while guaranteeing system safety invariants. (Supervisory Intervention)
Sim-to-Real Contact Exploit: An optimization pathology wherein a reinforcement learning policy discovers and exploits unphysical numerical artifacts in a simulator’s contact solver—such as interpenetration spring restitution, infinite friction sticking, or instantaneous impulse relaxation—to achieve high simulated reward. Because these mathematical solver relaxations do not exist in the physical world, the learned policy experiences catastrophic failure immediately upon encountering real Coulomb friction and non-interpenetrable material boundaries. Mitigating contact exploits requires non-smooth contact regularization, randomized contact parameters, and strict energy-passivity constraints during training. (Policy Training)
Sim-to-Real Gap (Sim-to-Reality Gap): The systematic discrepancy in dynamics, contact mechanics, sensor transduction, and execution latency that exists between a computational physics simulation and the real physical environment. Policies optimized purely in simulation frequently fail upon physical deployment because simulated rigid-body solvers linearize frictional contact, neglect transmission backlash and compliance, and omit high-frequency electrical bus latencies. Bridging this gap is the central challenge of simulation-based policy training in physical AI. (Policy Training, Closed-Loop Evaluation)
Spatial Information Age (\(t_{\text{age}}\)): The total wall-clock elapsed time between the physical capture epoch of a sensor observation and the exact moment that observation is utilized to make an actuation decision: \(t_{\text{age}} = t_{\text{act}} - t_{\text{capture}}\). In dynamic environments, nonzero information age artificially inflates the effective volume of physical obstacles by their maximum drift velocity (\(v_{\text{max}} t_{\text{age}}\)), penalizing slow inference pipelines with severe spatial clearance degradation. State estimation pipelines must enforce explicit information-age bounds to prevent policies from acting on stale world representations. (Sensor Perception, Spatial Memory)
Stopping Suffix: A pre-computed, kinodynamically verified braking trajectory permanently reserved at the terminal tail of an admitted action chunk, designed to bring the physical robot to a controlled, collision-free standstill within available friction and actuator limits. If real-time neural inference stalls, communication drops, or the subsequent action chunk fails kinematic validation, the robot smoothly transitions into the stopping suffix rather than executing open-loop extrapolation or triggering a jarring emergency brake. The stopping suffix ensures that loss of high-level intelligence never compromises low-level physical survival. (Trajectory Planning, Supervisory Intervention)

T

Takeover Reaction Budget (\(T_{\Sigma}\)): The total temporal allocation (\(T_{\Sigma} = t_{\text{detect}} + t_{\text{alert}} + t_{\text{cognitive}} + t_{\text{motor}} + t_{\text{plant}}\)) required for a human supervisor to detect a silent autonomy failure, evaluate the hazard, physically engage the controls, and command a dynamic maneuver to avert harm before the system enters an Inevitable Collision State. If the physical system’s time-to-harm is shorter than the minimum physiological and mechanical takeover budget, supervisory intervention is mathematically impossible. Systems operating in this regime must rely entirely on autonomous, real-time safety shielding rather than human fail-safe intervention. (Supervisory Intervention)
Temporal Ensembling: A real-time trajectory filtering technique for action-chunked policies that computes the final executed motor command as an exponentially weighted average of overlapping action predictions generated at successive observation steps. By continuously blending predictions produced across different historical contexts, temporal ensembling smooths inter-chunk discontinuities and eliminates high-frequency jerk without requiring explicit trajectory optimization. However, heavy ensembling windows introduce effective phase lag, reducing the robot’s peak reactivity during abrupt physical collisions or sudden obstacle appearances. (The Cognitive Brain, Policy Training, Trajectory Planning)
Time-to-Harm (\(t_{\text{harm}}\)): The minimum physical duration required for a system’s current kinetic state to evolve into an irreversible collision, structural failure, or human injury under worst-case unmodeled dynamics and disturbances. To preserve forward invariance, the permission path must continuously verify that \(t_{\text{harm}}\) remains strictly greater than the system’s worst-case reaction budget (\(t_{\text{detect}} + t_{\text{WCET}} + t_{\text{brake}}\)). If \(t_{\text{harm}}\) contracts below this reaction threshold, the machine has exhausted its decision margin and must initiate emergency braking. (The Epistemic Frontier)
Trajectory Spline Blending (\(C^2\) Jerk Continuity): A trajectory interpolation and seam-smoothing method that stitches consecutively admitted action chunks or trajectory segments using quintic polynomials or Hermite splines to enforce continuous position (\(C^0\)), velocity (\(C^1\)), and acceleration (\(C^2\)) across chunk boundaries. An unmanaged step change in acceleration requests an instantaneous step in motor torque (\(\tau = M \ddot{q}\)), implying infinite reflected jerk that excites mechanical drive compliance, damages transmissions, and triggers inverter overcurrent faults. Enforcing \(C^2\) continuity guarantees that commanded actuator efforts remain within dynamic bandwidth limits across receding-horizon replanning cycles. (Trajectory Planning, The Cognitive Brain)

V

Vision-Language-Action (VLA) Model: An end-to-end multimodal foundation model that consumes visual scene tokens, natural language task prompts, and robot proprioceptive states to directly generate continuous low-level action tokens or trajectory chunks. Built on high-capacity transformer backbones pretrained on Internet-scale vision-language datasets and fine-tuned on robotics demonstrations, VLA models synthesize open-vocabulary semantic reasoning with spatial affordance grounding. Because their internal weights lack physical safety and stability guarantees, their outputs are treated as unprivileged candidate proposals requiring downstream kinodynamic and barrier filtering. (The Cognitive Brain, Grounded Intent)

W

World Model: An internal generative model that learns to simulate the forward transition dynamics, contact mechanics, and sensory observations of the physical environment conditioned on the agent’s prospective actions. In physical AI, world models serve as learned digital twins that enable mental simulation, counterfactual trajectory evaluation, and model-predictive rollouts without inducing physical wear or hardware risk. Their reliability is fundamentally constrained by epistemic uncertainty over non-smooth contact bifurcations and the sim-to-real reality gap. (The Cognitive Brain, Trajectory Planning, Policy Training)