Reference desk
Guides
Durable explanations designed to make a fast-moving technical field understandable without sanding away uncertainty.
Contributed editorial
Beyond the Demo: How to Determine When a Humanoid Robot Is Actually Commercially Real
Randy Levine of Physical AI Radar explains how to separate humanoid demonstrations, pilots, contracts, deployments, and recurring commercial operations.
By Randy Levine, Physical AI Radar · Read editorial →
Humanoid Robots
Humanoid Robot Operating Hours: How to Read Runtime and Uptime Claims
A practical guide to comparing humanoid robot operating hours, runtime, uptime, utilization, intervention, and productive work without hiding fleet or schedule denominators.
Read guide →Robot Foundation Models
Robot Dataset Splits: What Does a Held-Out Test Actually Prove?
A practical audit of robot training, validation, and test splits, covering episodes, sites, objects, preprocessing, checkpoint selection, and generalization claims.
Read guide →Robot Learning
When Should a Robot Hand Control to a Human? An Intervention Evidence Guide
A practical robot intervention guide separating AutoIntervene's handoff evidence, operator time, physical recovery access, hazardous-energy isolation, and unassisted task success.
Read guide →Robot Foundation Models
Robot Benchmark Success Rates: A Practical Evidence Audit
A practical guide to reading robot benchmark results, using HumanCLAW, TurboVLA, SymmGrid, and ACT-2 to separate task scope, execution, trial counts, compute, and real-world evidence.
Read guide →Robot Learning
How Robot Training Data Is Made: Teleoperation, Egocentric Video, EEG, and EMG
A source-backed guide to robot training data, including real-robot, UMI, human video, simulation, and general data, plus labels, interventions, costs, and evidence quality.
Read guide →Embodied AI
Sim-to-Real Robot Transfer: Domain Randomization and the Hardware Evidence Gap
How robot policies move from simulation to hardware, what domain randomization can address, and how to test calibration, contact, latency, adaptation, and transfer.
Read guide →Humanoid Robots
Agility Robotics S-4: What Digit's Operating Hours, Pricing, Orders, and Losses Show
An evidence audit of Agility's SEC filing and Digit 5 launch, covering operating hours, pricing, order terms, losses, 2027 access, and missing denominators.
Read guide →Developer Tools
Physical AI Developer Stack: A Practical Map from ROS 2 to Robot Learning
A practical guide to choosing middleware, control, simulation, motion planning, data, policy training, evaluation, and deployment tools for a physical AI project.
Read guide →Embodied AI
What Is Embodied AI? A Practical Guide to Agents, Bodies, and Environments
A practical definition of embodied AI, how the perception and action loop works, how it differs from physical AI and robotics, and how to evaluate embodied agents.
Read guide →Foundations
What Is Physical AI?
A practical definition of physical AI, how the perception-to-action loop works, how it differs from embodied AI and robotics, and how to evaluate real systems.
Read guide →Developer Tools
Open-Source Robotics Project Health: A Practical Audit
A practical framework for evaluating ROS 2, Gazebo, MoveIt 2, LeRobot, Isaac Lab, and other open robotics projects by releases, support, governance, artifacts, tests, and reproducibility.
Read guide →Robot Foundation Models
What Is a Vision-Language-Action Model? How VLAs Control Robots
A practical guide to vision-language-action models, including inputs, action outputs, training data, architectures, evaluation, embodiment transfer, safety, and deployment limits.
Read guide →Humanoid Robots
Robot Demo Evidence Checklist: Autonomy, Resets, Speed, and Intervention
A practical checklist for judging robot videos and live demonstrations by task scope, continuity, autonomy, resets, speed, variation, failures, and operating evidence.
Read guide →Humanoid Robots
Boston Dynamics Atlas in 2026: Product, Pilot, and Deployment Evidence
A source-backed status audit of the product Atlas, including manufacturing, Hyundai and Google DeepMind allocations, specifications, control modes, customer access, and missing deployment results.
Read guide →Humanoid Robots
How Much Do Humanoid Robots Cost in 2026? Prices, Subscriptions, and Hidden Costs
A source-backed comparison of public humanoid robot prices and access models, including Unitree G1, 1X NEO, and Agility Digit, plus the costs that headline figures omit.
Read guide →Humanoid Robots
Figure at BMW: What Figure 02 Proved and Figure 03 Still Must Prove
A source-backed audit of Figure's BMW program, separating Figure 02's ten-month production record from Figure 03's logistics demonstration and next deployment phase.
Read guide →Humanoid Robots
Humanoid Robots in 2026: What Is Actually Available, Piloting, or Deployed?
An evidence-based comparison of leading humanoid robots, including Digit, Figure, XPENG IRON, AGIBOT G2, Atlas, Apollo, NEO, Phoenix, and Unitree, with manufacturing separated from sales and deployment.
Read guide →Embodied AI
Physical AI in Production: Six Deployments Compared by Evidence
A source-backed comparison of Gritt, Sereact, Monumental, Digit, Figure 02, and AGIBOT G2 across task, customer confirmation, fleet, runtime, throughput, intervention, safety, and commercial access.
Read guide →Robot Learning
LeRobot 0.6 Practical Guide: Models, Evaluation, and Human Correction
A source-backed guide to LeRobot v0.6.0, v0.6.1, and later main-branch changes, including models, evaluation, human correction, operator controls, cadence, and EO-1 language supervision.
Read guide →Humanoid Robots
Humanoid Robots as a Service: What a 90-Day Pilot Must Prove
A buyer's guide to humanoid Robots-as-a-Service, using Agility Robotics' deployment process to separate setup, pilot validation, production measurement, support, safety, and ROI evidence.
Read guide →Robot Foundation Models
World Action Models for Robot Control: What Future Prediction Must Prove
An evidence guide to robot world models, including action compliance, compact latent intentions, control latency, physical trials, and simulation limits.
Read guide →Embodied AI
Maven Robotics Evidence Audit: Deployment, Uptime, and the Missing Denominators
A source-backed audit of Maven Robotics, covering its $100 million Series A, mixed-case palletizing, reported robot fleet, 16-hour days, 99% uptime, expansion plans, and missing operating evidence.
Read guide →Embodied AI
Physical AI Funding Tracker: What Has the Capital Actually Produced?
A source-backed comparison of Unitree, Atoms, Humanoid, Walden, Apptronik, Agility Robotics, Gritt, and Sereact by capital, manufacturing, customers, and operating evidence.
Read guide →Robot Learning
FolDeX Evidence Audit: What 2,000 Hours of Real-Robot Data Actually Prove
A source-backed audit of FolDeX and FoldChallenge, covering real-robot data scale, recovery trajectories, transfer, held-out garments, physical evaluation, FoldScore, access, and evidence limits.
Read guide →Robot Foundation Models
When Robot Rewards Change With Wording: Auditing Vision-Language Reward Models
A practical audit of paraphrase fragility in robot reward models, including score crossing, flip rate, selection regret, mitigation, and deployment limits.
Read guide →Humanoid Robots
When Is an Open-Source Humanoid Actually Reproducible? A Bridge Evidence Audit
A practical audit of Bridge and the evidence needed to reproduce an open-source humanoid across hardware, control, cost, safety, and physical testing.
Read guide →Robot Foundation Models
The Embodiment Gap in Robot Foundation Models: What Still Changes on the Target Robot
A practical guide to cross-embodiment robot claims, including strict versus pretrain-exposed zero-shot transfer, target data, calibration, control, safety, recovery, and evaluation.
Read guide →Robot Foundation Models
Contact-Rich Robot Manipulation: How to Audit Force, Precision, Adaptation, and Recovery
A practical evidence guide to force-aware assembly, insertion tolerance, contact quality, adaptation cost, recovery, and reproducible physical evaluation.
Read guide →Robot Foundation Models
Are Physical AI Benchmarks Measuring the Same Thing? A Redundancy and Ranking Audit
A practical audit of benchmark overlap, rank sensitivity, general capability, compact suites, and what current physical AI leaderboards do not establish.
Read guide →Robot Foundation Models
Why One Robot Success Rate Hides the Failure: A Diagnostic Benchmark Guide
A practical guide to observation ablations, skill-level evaluation, restorable intermediate states, failure coverage, and the evidence hidden by one robot success rate.
Read guide →Robot Safety
When Robots Mishear: How to Audit Voice Command Safety
An evidence guide to speech recognition errors in robot control, including intent confirmation, command authority, safe refusal, execution gates, testing, and field metrics.
Read guide →Robot Foundation Models
Can a Robot Learn in Five Minutes? What Rapid On-Robot Adaptation Really Costs
An evidence guide to physical interaction time, prior knowledge, resets, human support, compute, safety constraints, and retention in rapid on-robot learning.
Read guide →Robot Foundation Models
VLA Control Interfaces: Prompt Authority, Spatial Grounding, Intent, and Memory
A practical evidence guide to four interfaces that shape vision-language-action control: prompt authority, typed spatial grounding, behavior intent, and event-selected memory.
Read guide →Robot Foundation Models
Can Robot Policies Learn from Failure? What Q-Planning Actually Proves
An evidence audit of Q-Planning, which uses a learned value model to improve action selection while keeping a large behavior-cloning robot policy frozen.
Read guide →Robot Foundation Models
Latent Actions for Robot Learning: What 41 Design Choices Actually Show
A practical evidence guide to latent action models, proxy metrics, action integration, data scaling, and the limits of a 59-million-frame robot learning study.
Read guide →Robot Safety
Adaptive Robot Safety Cases: Specifications, Monitoring, and Evidence
A practical guide to building an evidence-backed safety case for an adaptive robot, using ManiGuard to separate task success, safe success, specifications, runtime monitoring, distribution shift, and field validation.
Read guide →Robot Foundation Models
Long-Horizon Robot Control: Planning, Memory, and Whole-Body Execution
An evidence guide to long-horizon robot control, using Tau-0-VLA and HAF to separate subtask planning, execution memory, low-level action, whole-body coordination, and online adaptation.
Read guide →Robot Foundation Models
Dynamic Robot Benchmarks: How to Measure Latency, Recovery, and Process Quality
A practical evidence guide to dynamic robot evaluation, using ReflexBench and PRM-as-a-Judge 1.5 to measure reaction timing, progress, regression, recovery, and execution quality.
Read guide →Robot Foundation Models
Xiaomi-Robotics-1: What 100,000 Hours of Training Data Actually Prove
A source-backed audit of Xiaomi-Robotics-1, including its 100,000-hour UMI corpus, released code and 5B checkpoint, benchmark protocols, and missing evidence.
Read guide →Robot Learning
Ego2Robot: What 18,561 Hours of Synthetic Robot Data Actually Mean
An evidence audit of Ego2Robot, covering its 1,940 source hours, 15 rendered morphologies, quality filters, fixed-frame comparisons, real-robot tests, artifacts, and limits.
Read guide →Robot Foundation Models
Gemini Robotics 2: What the Whole-Body, Reasoning, and On-Device Models Actually Prove
An evidence audit of Google DeepMind's three Gemini Robotics 2 models, covering whole-body control, embodied reasoning, on-device adaptation, access, safety, and missing deployment evidence.
Read guide →Robot Policy
FCC Foreign-Robot Restrictions: What the Covered List Changes for U.S. Buyers
A source-backed guide to the FCC's July 2026 restrictions on foreign-produced mobile robots, including scope, existing models, Conditional Approvals, software updates, and buyer due diligence.
Read guide →Robot Safety
Robot Safety Standards for Working Near People
A source-backed guide to ISO 10218, ISO 13482, ISO 3691-4, collaborative robot applications, risk assessment, and choosing the right safety framework for industrial robots, AMRs, service robots, and humanoids.
Read guide →Robot Foundation Models
FLUX-mimic: What the Video-Action Model and Audi Evidence Actually Prove
A source-backed audit of FLUX-mimic, including its FLUX 3 video backbone, action decoder, 20-trial benchmark, edge latency, Audi work, model access, and missing operating evidence.
Read guide →Robot Foundation Models
Sunday Robotics ACT-2: What 778 of 785 Laundry Folds Actually Prove
A source-backed audit of ACT-2’s reported 99.1% laundry-folding result, including scope, adaptation cost, grading, speed, generalization claims, and the evidence still needed from home beta testing.
Read guide →Industrial Robots
Walden Robotics Launch: What the Toyota Factory Evidence Actually Proves
A source-backed analysis of Walden Robotics’ $300 million launch, Toyota factory operation, wheeled robot, AI stack, commercial access, and the deployment evidence still missing.
Read guide →