Physical AI Foundations

Video (best)

• MIT OpenCourseWare / Russ Tedrake — “Underactuated Robotics Lecture Series (Perception & Control)”
• Link: https://ocw.mit.edu/courses/6-832-underactuated-robotics-spring-2009/
• Why: Russ Tedrake’s MIT 6.832 lectures are the closest rigorous academic treatment of the perception-action loop, embodied intelligence, and autonomous systems foundations. However, no single YouTube video cleanly covers “Physical AI Foundations” as a unified topic with the sensor fusion + embodied intelligence framing used in modern Physical AI discourse.
• Level: advanced

⚠️ Coverage Gap Note: No single YouTube explainer from the preferred educators (3Blue1Brown, Karpathy, Yannic, StatQuest) covers Physical AI Foundations as a unified concept. The term “Physical AI” as used by NVIDIA/industry (convergence of perception, action, and embodied intelligence) is too recent (2023–2024) for polished explainers to exist.

Best available alternative:

• NVIDIA GTC Keynote clips — Jensen Huang’s GTC 2024 “Physical AI” segment on YouTube provides the conceptual framing, but is a product keynote, not a pedagogical resource. [Video ID: _waPvOwL9Z8 (GTC March 2025 Keynote)]

---

Blog / Written explainer (best)

• Lilian Weng — “Autonomous Agents” (covers embodied intelligence, perception-action loops, and agent grounding)
• Link: https://lilianweng.github.io/posts/2023-06-23-agent/
• Why: Weng’s characteristic depth and clarity covers the cognitive architecture underlying Physical AI — perception, memory, action, and tool use — with rigorous citations. While not exclusively about robotics hardware, it is the best written explainer from a trusted educator that addresses the embodied intelligence and perception-action loop concepts central to this topic. Her treatment bridges the gap between LLM-based reasoning and physical grounding.
• Level: intermediate/advanced

---

Deep dive

• Author/Source — Russ Tedrake, “Robotic Manipulation: Perception, Planning, and Control” (MIT Open Textbook)
• Link: https://manipulation.csail.mit.edu/
• Why: This is the most comprehensive freely available technical reference covering the full stack of Physical AI: depth cameras, sensor fusion, perception-action loops, grasping, and manipulation planning. Written by one of the field’s leading researchers, it includes code, lecture notes, and problem sets. Covers sim-to-real, tactile sensing concepts, and safety considerations. Far more rigorous than any blog post for learners who need implementation depth.
• Level: advanced

---

Original paper

• None identified — Physical AI as a unified concept does not have a single seminal “founding paper” in the way that, e.g., the Transformer or AlexNet do. The topic is a convergence of multiple subfields.

Closest candidates (use with caveat):

• For embodied intelligence: Pfeifer & Bongard’s work, or the RT-2 paper (arxiv.org/abs/2307.15818) for vision-language-action models
• For sensor fusion foundations: No single canonical paper
• For tactile sensing (GelSight): Yuan et al., 2017 — “GelSight: High-Resolution Robot Tactile Sensors for Estimating Geometry and Force” https://pmc.ncbi.nlm.nih.gov/articles/PMC5751610/

Recommendation: Do not force a single paper here. Assign RT-2 + one tactile sensing paper as a paired reading instead.

---

Code walkthrough

• Source — MIT Manipulation Course Notebooks (Drake-based)
• Link: https://manipulation.csail.mit.edu/clutter.html
• Why: Tedrake’s course provides runnable Jupyter/Colab notebooks using the Drake simulator that walk through perception pipelines (depth cameras, point clouds), grasp planning, and sensor integration — directly mapping to the core concepts of Physical AI Foundations. These are the most pedagogically complete hands-on implementations available for this topic from a credible academic source.
• Level: advanced

Alternative for beginners:

• NVIDIA Isaac Sim tutorials cover sensor fusion and robot perception in a more accessible way, but require proprietary software setup. [NOT VERIFIED]

---

Coverage notes

• Strong: Perception-action loop, embodied intelligence (conceptual), manipulation and depth cameras (Tedrake’s materials), autonomous systems architecture
• Weak: Tactile sensing (GelSight) — no strong standalone tutorial exists outside research lab pages; LiDAR-specific educational content tends to live in autonomous driving courses, not general Physical AI courses
• Weak: Safety certification for physical AI systems — this is a genuine educational gap; most content is either too theoretical (formal verification literature) or too domain-specific (automotive ISO 26262)
• Gap: No high-quality beginner-to-intermediate YouTube series exists for Physical AI Foundations as a unified topic. This is a significant gap for the platform — original content creation would add real value here.
• Gap: Sensor fusion for humanoid robots specifically (as opposed to autonomous vehicles) has almost no dedicated educational resources outside of proprietary robot SDK documentation.

---

---

Additional Resources for Tutor Depth

10 sources — papers, official docs, working code, benchmarks, and deep explainers that give the AI tutor precision on this topic.

📄 Diffusion Policy (Action Diffusion for Visuomotor BC)

Paper · source

Full diffusion-policy objective + sampling procedure; benchmark tables across multiple manipulation suites

Key content

• DDPM sampling / policy inference (Eq. 1, Eq. 4): start from Gaussian noise (x_K) (or action-seq (A_t^K)). Iterate
  (x_{k-1}=\alpha\big(x_k-\gamma,\varepsilon_\theta(x_k,k)+\mathcal N(0,\sigma^2 I)\big)) (Eq. 1).
  Conditional visuomotor form: (A_{t}^{k-1}=\alpha\big(A_t^k-\gamma,\varepsilon_\theta(O_t,A_t^k,k)+\mathcal N(0,\sigma^2 I)\big)) (Eq. 4).
  Interpretable as noisy gradient step (x’ = x-\gamma\nabla E(x)) (Eq. 2).
• Training objective (Eq. 3, Eq. 5): sample data (x_0) (or (A_t^0)), pick random diffusion step (k), add noise (\varepsilon_k). Minimize
  (L=\mathrm{MSE}(\varepsilon_k,\varepsilon_\theta(x_0+\varepsilon_k,k))) (Eq. 3); conditional:
  (L=\mathrm{MSE}(\varepsilon_k,\varepsilon_\theta(O_t,A_t^0+\varepsilon_k,k))) (Eq. 5).
• Closed-loop receding horizon (Sec. 2.3): observe last (T_o) steps (O_t); predict (T_p) future actions; execute only (T_a) then replan; warm-start next plan with unexecuted actions.
• Design defaults: Square Cosine noise schedule (Sec. 3.3). DDIM for speed: 100 training iterations, 10 inference iterations → ~0.1s latency on RTX 3080 (Sec. 3.4). Action horizon 8 steps often best (Sec. 5.3, Fig. 5).
• Vision encoder (Sec. 3.2): ResNet-18 (no pretrain), spatial softmax (replace GAP), GroupNorm (replace BatchNorm) for EMA stability.
• Empirical headline: across 15 tasks / 4 benchmarks, Diffusion Policy improves average success by 46.9% vs prior SOTA (Abstract/Sec. 5.3).
• Key benchmark numbers:
  • RoboMimic visual (Table 2): Transport (PH) LSTM-GMM 0.88/0.62 vs DiffusionPolicy-C 1.00/0.93; ToolHang (PH) 0.68/0.49 vs 0.95/0.73.
  • Multi-stage state (Table 4): BlockPush p2 BET 0.71 vs DiffusionPolicy-T 0.94; Kitchen p4 BET 0.44 vs DiffusionPolicy-C 0.99 (T: 0.96).
  • Real Push-T (Table 6): Diffusion Policy (end2end) Succ 0.95, IoU 0.80; IBC 0.00, LSTM-GMM best 0.20.

📄 Diffusion Policy (Action Diffusion for Visuomotor Control)

Paper · source

Benchmark success rates across manipulation suites + diffusion-policy objective & action-sequence sampling (DDPM/DDIM)

Key content

• Policy as conditional DDPM over action sequences (Sec. 2): generate an action trajectory by iterative denoising, conditioned on observation (o).
  • Sampling update (Eq. 1 / Eq. 4): (a_{t-1} = f_\theta(a_t, t, o) + \sigma_t z), with (z\sim\mathcal N(0,I)). (Unconditional form in Eq. 1; conditional on (o) in Eq. 4.)
  • Gradient-descent view (Eq. 2): update interpretable as a noisy gradient step where the network predicts a score/gradient field; (\sigma_t) controls noise, step size acts like a learning rate.
• Training objective (Eq. 3 / Eq. 5): sample data action sequence (a_0), choose diffusion step (t), add Gaussian noise to get (a_t); train (\epsilon_\theta(\cdot)) to predict the injected noise (MSE on noise prediction), conditioned on (o).
• Closed-loop receding-horizon execution (Sec. 2.3): at time (k), input last (T_o) observations, predict (T_a) future actions, execute first (T_e) actions, then replan; can warm-start next plan from previous predicted sequence.
• Real-time acceleration (Sec. 3.4): use DDIM; example: 100 training diffusion steps, 10 inference steps → ~0.1 s latency on NVIDIA 3080.
• Design choices/rationale
  • Condition on observations (not joint obs-action diffusion) to avoid inferring future states → faster inference, feasible end-to-end vision training (Sec. 2.3).
  • CNN vs Transformer (Sec. 3.1): CNN works “out of the box”; Transformer helps when actions change rapidly (reduces CNN over-smoothing) but is more hyperparameter-sensitive.
  • Noise schedule: Square Cosine (iDDPM) worked best empirically (Sec. 3.3).
• Key empirical results (success rates)
  • Across 15 tasks / 4 benchmarks, Diffusion Policy improves average success by 46.9% (Abstract/Sec. 5).
  • BlockPush / Kitchen table: DiffusionPolicy-T: BlockPush p1=0.99, p2=0.94; Kitchen p1=1.00, p2=0.99, p3=0.99, p4=0.96 (beats BET p4=0.44).
  • Real-world Push-T: Diffusion Policy (E2E vision) Succ%=0.95, IoU=0.80 vs IBC best Succ%=0.00 and LSTM-GMM best Succ%=0.20 (Tab. 6).
  • Real-world Pour/Spread: Diffusion Policy Pour IoU=0.74, Succ=0.79; Spread Coverage=0.77, Succ=1.00 vs LSTM-GMM Succ=0.00 both (Sec. 6.3 table).

📄 Diffusion Policy (DDPM for visuomotor action sequences)

Paper · source

Proceedings version with definitive method description, experimental protocol, and results tables

Key content

• DDPM denoising update (Eq. 1):
  (x_{k-1}=\alpha\big(x_k-\gamma,\epsilon_\theta(x_k,k)+\mathcal{N}(0,\sigma^2 I)\big)).
  Interpretable as noisy gradient step (Eq. 2): (x’ = x-\gamma\nabla E(x)), where (\epsilon_\theta) predicts the gradient field.
• Training objective (Eq. 3): sample data (x_0), pick random step (k), add noise (\epsilon_k); minimize
  (L=\mathrm{MSE}(\epsilon_k,\epsilon_\theta(x_0+\epsilon_k,k))).
• Policy conditioning + closed-loop horizons (Sec. II-C, Fig. 3): learn conditional (p(A_t|O_t)) with action sequences.
  Conditioned denoising (Eq. 4): (A^{k-1}t=\alpha(A^k_t-\gamma,\epsilon\theta(O_t,A^k_t,k)+\mathcal{N}(0,\sigma^2I))).
  Loss (Eq. 5): (L=\mathrm{MSE}(\epsilon_k,\epsilon_\theta(O_t,A^0_t+\epsilon_k,k))).
  Horizons: observation (T_o), prediction (T_p), execute (T_a) steps before replanning (receding horizon).
• Architectures (Sec. III-A): CNN backbone with FiLM conditioning; Time-series Diffusion Transformer (cross-attention to obs; causal attention over actions) reduces CNN over-smoothing for high-rate control.
• Visual encoder (Sec. III-B): ResNet-18 (no pretrain), spatial softmax pooling, GroupNorm (stable with EMA).
• Defaults/acceleration: Square Cosine noise schedule (iDDPM) works best (Sec. III-C). DDIM speeds inference: 100 training steps, 10 inference steps → ~0.1s latency on RTX 3080 (Sec. III-D).
• Key empirical results:
  • Across 12 tasks / 4 benchmarks, Diffusion Policy average +46.9% success-rate improvement vs prior SOTA (Abstract/Sec. V).
  • Real Push-T: Diffusion Policy (end-to-end transformer) 95% success, IoU 0.80 vs Human 1.00 / 0.84; IBC best 0%, LSTM-GMM best 20% (Table V).
  • Mug flip: Diffusion Policy 90% vs LSTM-GMM 0% (Fig. 10).
  • Sauce tasks: Pour IoU 0.74, succ 0.79; Spread coverage 0.77, succ 1.00 (Fig. 11).
• Design rationale: conditioning on (O_t) (not joint (A,O)) extracts vision once → real-time control; action-sequence diffusion handles multimodality + temporal consistency; score modeling avoids EBM normalization/negative sampling → training stability (Sec. IV-D).

📄 RT-2 (Vision-Language-Action) — tokenized actions + co-fine-tuning boosts generalization

Paper · source

RT-2 evaluation results (success rates across seen/unseen + emergent skills) and ablations (model size, co-fine-tuning vs robot-only vs scratch)

Key content

• Core idea (Sec. 3.2): Train a pretrained VLM to output robot actions as text tokens (Vision-Language-Action / VLA).
  • Action space: 6-DoF end-effector Δposition (x,y,z) + Δrotation (x,y,z) + gripper extension + terminate.
  • Discretization: all continuous dims discretized into 256 uniform bins → action represented as 8 integer tokens.
  • Output string format: "terminate Δposx Δposy Δposz Δrotx Δroty Δrotz gripper" (example token sequence shown in paper).
• Token mapping:
  • PaLI-X: integers up to 1000 have unique tokens → map bins directly to integer tokens.
  • PaLM-E: overwrite 256 least-frequent tokens to serve as action vocabulary (“symbol tuning”).
• Training procedure (Sec. 3.2): Co-fine-tune on (a) robot trajectories + (b) original web VLM mixture (VQA/captioning/interleaved image-text). Increase robot sampling weight to balance batches.
• Decoding constraint: when prompted for robot action, restrict output vocabulary to valid action tokens only.
• Real-time control (Sec. 3.3): serve models via cloud TPU; RT-2-PaLI-X 55B runs ~1–3 Hz, 5B runs ~5 Hz.
• Empirical results:
  • ~6,000 real-robot evaluation trials total.
  • Generalization: RT-2 models achieve ~2× average improvement over next best baselines (RT-1, MOO) on unseen objects/backgrounds/environments; ~6× over other baselines (Sec. 4.1, Fig. 4).
  • Emergent skills: best RT-2 achieves >3× average success vs next best baseline (RT-1) across symbol understanding/reasoning/human recognition (Sec. 4.2, Fig. 6a).
  • Ablations (Sec. 4.3, Fig. 6b): training from scratch performs poorly; co-fine-tuning > robot-only fine-tuning; 55B > 5B for generalization.
• Language-Table sim benchmark (Table 1): RT-2-PaLI-3B 90 ± 10 vs BC-Zero 72 ± 3, RT-1 74 ± 13, LAVA 77 ± 4.

📄 Robustness of LiDAR vs RADAR vs Depth Camera on ill-reflective surfaces

Paper · source

Measured robustness comparisons across sensing modalities (LiDAR vs depth camera vs radar) with failure modes and performance under adverse (ill-reflective) conditions

Key content

• Sensor principles (Section II):
  • LiDAR: active ToF in near-infrared; detects returned reflected light along emitted beam path.
  • FMCW RADAR: range from beat frequency; range resolution (\Delta R = \frac{c}{2B}). Velocity from phase differences across chirps; AoA via beamforming + CFAR to form point clouds.
  • Depth camera (Intel D455): IR stereo; depth from disparity: Eq. 1 (Z = \frac{f,b}{d}) where (Z)=depth, (f)=focal length, (b)=baseline, (d)=disparity.
• Evaluation metrics (Section IV):
  • Eq. 2 (MAE): (\text{MAE}=\frac{1}{N}\sum_{i=1}^{N}\left|r^{gt}_i-r^{meas}_i\right|).
  • Eq. 3 (Std): (\sigma=\sqrt{\frac{1}{N}\sum_{i=1}^{N}(x_i-\mu)^2}).
• Material reflectivity (Section V-A, Table I):
  • White matt max power reflectivity 9.5% vs black matt 4.7% (s-pol) and 0.3% (p-pol); black p-pol reflectivity 13× lower than black s-pol.
• Static ranging results (Section V-B):
  • Baseline (white): LiDAR max 10 m; MAE <5 cm up to 8.5 m. RADAR max 7 m, MAE ~10–15 cm. Depth camera max 6 m, MAE 61 cm at 6 m (monotonically worsens with distance).
  • Ill-reflective (black): LiDAR max 3 m (33% of baseline). RADAR max 7.5 m, MAE <10 cm for ≥2.5 m but more false positives at short range. Depth camera keeps 6 m max range; similar trend, lower MAE beyond 3 m vs baseline.
• Dynamic downstream impact (Section III-C, V-C):
  • Track half reflective/half ill-reflective; 10 laps each config; fusion prioritizes LiDAR, uses depth/RADAR when LiDAR unavailable.
  • Mean lap times: LiDAR+Depth 7.686 s (best); LiDAR-only 7.729 s (+0.043 s, 0.573% worse); LiDAR+RADAR 8.011 s (12.026% worse vs LiDAR+Depth). Std dev ≤ 0.054 s.

📊 Waymo public-road safety metrics & benchmarks

Benchmark · source

Operational safety metrics + comparative crash/incident rates for public-road autonomous operation

Key content

• Exposure (Abstract): >6.1M miles automated driving in Phoenix ODD; includes 65,000 miles driverless (2019–first 9 months 2020).
• Event accounting (Section 3): 47 total contact events = 18 actual + 29 simulated (from counterfactual “what-if” post-disengagement simulation). 1 event during driverless operation.
• Severity framework (Section 2.3 / Table 1 description): Events categorized using ISO 26262 severity classes S0–S3 (S0 no injury expected → S3 possible critical injuries). Waymo estimates injury likelihood using AIS-linked probabilities; severity estimation uses impact object, impact velocity, impact geometry, and ΔV / principle direction of force.
• Airbag-deployment-level events (Table 1 notes): 8 most severe events defined as involving actual or expected airbag deployment; all 8 involved road rule violations/other errors by human road users. No actual or predicted S2 or S3 events.
• Counterfactual simulation workflow (Section 2.3):
  1. Re-simulate AV post-disengagement using recorded pre-disengage state (position/attitude/velocity/acceleration) + recorded sensor observations to generate AV’s counterfactual trajectory.
  2. If needed, simulate other agents using deterministic human collision-avoidance behavior models (calibrated on naturalistic collision/near-collision data; response triggered by deviations from expectations).
  3. Infer contact; assign severity via national crash databases; add scenario to regression library.
• Collision-mode result (Discussion): 0 actual or simulated events in “road departure, fixed object, rollover” typology (noted as 27% of US roadway fatalities).
• Benchmarking rationale (Comparison benchmarks): Comparisons to human crash stats are hard due to ODD specificity, different crash definitions (Waymo includes minor contacts), and model assumptions; use rates per mile and confidence bounds (6.1M miles supports S0/S1 signal, not rare S2/S3).

📖 Bayes Filter → Kalman/EKF foundations

Reference Doc · source

Bayes filter framing; Markov assumptions; KF/EKF live in Ch.3 (Gaussian filters)

Key content

• Belief (posterior) definition (Eq. 2.35):
  [ bel(x_t)=p(x_t\mid z_{1:t},u_{1:t}) ] where (x_t)=state at time (t), (z_t)=measurement, (u_t)=control.
• Prediction belief (Eq. 2.36):
  [ \overline{bel}(x_t)=p(x_t\mid z_{1:t-1},u_{1:t}) ]
• Markov/conditional independence assumptions:
  State transition (Eq. 2.33):
  [ p(x_t\mid x_{0:t-1},z_{1:t-1},u_{1:t})=p(x_t\mid x_{t-1},u_t) ] Measurement model (Eq. 2.34):
  [ p(z_t\mid x_{0:t},z_{1:t-1},u_{1:t})=p(z_t\mid x_t) ] These define a DBN/HMM with transition (p(x_t\mid x_{t-1},u_t)) and sensor model (p(z_t\mid x_t)) plus initial (p(x_0)).
• Bayes rule (Eq. 2.15):
  [ p(x\mid y)=\eta,p(y\mid x)p(x) ] ((\eta)=normalizer independent of (x)).
• Total probability (Eq. 2.11–2.12):
  [ p(x)=\sum_y p(x\mid y)p(y)\quad\text{or}\quad p(x)=\int p(x\mid y)p(y),dy ]
• Gaussian PDF (Eq. 2.4):
  [ \mathcal N(x;\mu,\Sigma)=\det(2\pi\Sigma)^{-1/2}\exp{-\tfrac12(x-\mu)^T\Sigma^{-1}(x-\mu)} ] (basis for KF/EKF in Chapter 3).

📖 ROS 2 QoS Policies + Standard Profiles (Iron)

Reference Doc · source

ROS 2 QoS policy definitions + predefined QoS profiles (defaults for pub/sub, services, sensor data, parameters) and compatibility rules.

Key content

• QoS = profile of policies applied per publisher/subscription/service client/server; incompatible QoS can prevent any message delivery.
• Core QoS policies (definitions):
  • History: Keep last (store up to N samples) vs Keep all (store all, limited by middleware resources).
  • Depth: queue size N, only honored if History=Keep last.
  • Reliability: Best effort (may drop) vs Reliable (guaranteed delivery; may retry).
  • Durability: Transient local (publisher persists samples for late joiners) vs Volatile (no persistence).
  • Deadline: expected max time between publishes.
  • Lifespan: max time from publish to receive before message is stale; expired messages silently dropped.
  • Liveliness: Automatic vs Manual by topic (publisher must assert alive via API).
  • Lease duration: max time allowed without liveliness assertion before considered lost.
• Default pub/sub profile (ROS 1–like): History=Keep last, Depth=10, Reliability=Reliable, Durability=Volatile, Liveliness=System default; Deadline/Lifespan/Lease duration = Default (unspecified).
• Services profile rationale: Reliable + Volatile (avoid restarted servers receiving outdated requests; client protected from multiple responses, server not protected from side-effects).
• Sensor data profile rationale: prioritize timeliness over completeness → Best effort + smaller queue (than default).
• Parameters profile: like services but much larger depth to avoid losing requests when client can’t reach server.
• Compatibility model (Request vs Offered): subscriber requests “minimum acceptable”; publisher offers “maximum provided”; connect only if every policy requested is not more stringent than offered.
  • Reliability compatibility (Pub→Sub): BE→BE Yes; BE→Reliable No; Reliable→BE Yes; Reliable→Reliable Yes.
  • Durability compatibility: Volatile→Transient local No; Transient local→Transient local Yes (new+old); Transient local→Volatile Yes (new only). Latched behavior requires both sides Transient local.
  • Deadline/Lease duration: Default→x No; x→Default Yes; x→y compatible iff y ≥ x.
  • Liveliness: Automatic→Manual No; Manual→Automatic Yes.
• QoS events: offered/requested deadline missed, liveliness lost/changed, offered/requested incompatible QoS; plus matched events on connect/disconnect.

📋 # Source: https://github.com/TheGreatGalaxy/sensor-fusion

Source ·

🔍 Waymo Safety Case (AUR) structure + evidence credibility

Explainer · source

Safety-case structure: hazard/risk framing, safety argumentation, evidence types (simulation + on-road) for deployment

Key content

• Safety case definition (UL 4600:2022): “Structured argument, supported by a body of evidence… that a system is safe for a given application in a given environment.”
• Top-level goal: Absence of Unreasonable Risk (AUR); treated as a risk assessment problem requiring explicit Acceptance Criteria (AC) + Validation Targets (ISO 21448:2022; ISO/AWI TS 5083).
• Key formula (risk):
  Risk = P(harm) × Severity(harm) (ISO 26262:2018 framing).
  Variables: P(harm) probability of harm; Severity(harm) magnitude of harm.
• Hazard decomposition (Section 2.2): pin residual risk to 3 hazard categories:
  1. Architectural (e.g., sensor placement blindspots)
  2. Behavioral (e.g., unsafe proximity / driving behavior)
  3. In-service operational (e.g., cargo securement, malicious access)
• Causal chain framing (Section 2.2.1): scenario triggering conditions → hazardous causal element/behavior → hazard manifestation → harm; risk influenced by exposure + controllability (ISO concepts).
• Behavioral AC framework = 5D coverage space (Section 2.3):
  1. Severity potential (injury severity; AIS scale 1–6)
  2. Conflict role (ADS as initiator vs responder)
  3. Behavioral capability: regulatory compliance, conflict avoidance, collision avoidance
  4. ADS functionality status: nominal vs degraded
  5. Level of aggregation: event-level + aggregate-rate ACs (use both; aggregates alone can miss rare scenario risks)
• Safety determination lifecycle (Section 3.1): (1) safety-by-design development → (2) readiness review using on-road data (10+ years) + simulation + expert judgment; failures to meet targets delay approval → (3) post-deploy monitoring for continuous confidence growth.
• Credibility structuring (Section 4 overview): Case Credibility Assessment (CCA) = credibility of argument (top-down) + credibility of evidence (bottom-up) + implementation credibility check.

---

Related Topics

• Physics Simulation
• Robot Manipulation
• Reinforcement Learning