comparisons
Event Cameras vs Frame Cameras
A frame camera blinks 30 times a second; an event camera reports light changes in microseconds. Here is when that speed earns its cost in robot capture.
6 min read
JOURNAL
Notes on Physical AI.
How humanoid robots learn, where their training data comes from, and why the data layer is becoming the real bottleneck. Clear, sourced writing for engineers, operators, and investors.
comparisons
Onboard vs Cloud Robot Inference
Onboard robot inference caps the control rate you can carry; cloud inference trades latency for scale. How that choice quietly shapes what you must capture.
5 min read
comparisons
Single-Arm vs Bimanual Manipulation Data
Why bimanual manipulation data costs far more than twice single-arm data, the coordination tax, and what two coordinated hands demand of capture.
5 min read
comparisons
Motion Capture Suit vs Vision-Only Capture
A motion capture suit nails fingertip accuracy, markerless vision scales in the wild. Here is the real cost and fidelity tradeoff for human capture.
5 min read
comparisons
Wrist Camera vs Third-Person View
A wrist camera sees the fingertips, a third-person view sees the room. Here is what each teaches a manipulation policy, and why you capture both.
5 min read
comparisons
RGB vs Depth Cameras for Robots
When a depth camera earns its bandwidth and when plain RGB is enough for robot manipulation, and what to capture so you never lose the geometry.
6 min read
foundations
Self-Supervised Learning for Robotics
Self-supervised learning lets robots learn from unlabeled interaction instead of costly labels, and why first-person egocentric video is the natural fit.
5 min read
foundations
Offline Reinforcement Learning for Robots
Offline reinforcement learning turns logged robot data into a policy without a simulator. Its real promise, its one hard failure mode, and its data appetite.
5 min read
foundations
Representation Learning for Manipulation
Two policies, same demos, different features, opposite results. How representation learning quietly decides whether a manipulation policy generalizes.
5 min read
foundations
Task Horizon: Why Long Tasks Get Exponentially Harder
Chain thirty reliable steps and the task fails four times in five. The task horizon problem, why long-horizon errors compound, and the data that fixes it.
5 min read
foundations
What Counts as a Good Demonstration?
Not every demonstration teaches a robot. The quality dimensions of a single demo, why bad ones poison imitation learning, and how they compound.
5 min read
foundations
Episodic Data vs Streaming Robot Trajectories
Episodic data or a continuous stream? The framing choice behind every robot trajectory dataset, and how it decides what a policy can actually learn.
5 min read
foundations
Goal-Conditioned Policies: One Model, Many Goals
A goal-conditioned policy takes the goal as input, so one model chases many goals. Here is the paired data, and the hindsight trick, that make it work.
5 min read
foundations
Reward Models for Robot Learning, and Why Rewards Fail
Hand-designed rewards break in real-world manipulation. Here is why, and how a learned reward-model approach from preferences, demos, and video replaces them.
5 min read
foundations
Proprioception: The Sense Robots Cannot See
Proprioception, a robot's sense of its own joints and forces, is what manipulation relies on when vision is blocked, and why passive video cannot supply it.
5 min read
foundations
Observation and Action Spaces: The Quiet Decisions
The action space and observation space you pick decide what a dataset can ever teach. A field guide to the quiet choices that outlast the robot.
5 min read
foundations
What Is a Robot Policy? Observation to Action
A robot policy maps observation to action. See how output rate, action format, memory, and determinism decide exactly what data you must capture.
5 min read
foundations
Behavior Cloning Explained: The Default, and Its Limits
Behavior cloning is the workhorse of robot learning. Why it became the default, the compounding-error failure that breaks it, and what better data fixes.
5 min read
policy
Why Data Residency and Data Sovereignty Are Not the Same
Discover why the precise legal distinction between residency and sovereignty changes how EU buyers purchase human-demonstration robot data.
6 min read
deep-dives
Why Domain Randomization is the Workhorse of Sim-to-Real
Discover how domain randomization bridges the simulation-to-reality gap in robotics, where it fails, and why real-world data remains the ultimate bottleneck.
6 min read
foundations
Skill Primitives: Building Complex Tasks from Reusable Parts
How modular policies and skill primitives solve the long-horizon task problem in robotics, and why our data strategy must change to support them.
5 min read
deep-dives
Why Force Feedback is the Missing Signal in Teleoperation
Visual-only imitation learning fails at the contact boundary. We explore how active force feedback transforms both operator performance and physical AI data.
5 min read
foundations
Hand-Pose Estimation: Reading Fingers from First-Person Video
How extracting 21-keypoint coordinates from egocentric video solves the human-to-robot hand retargeting problem for dexterous manipulation.
5 min read
trends
Physical AI in Europe: the map, and the missing layer
Europe's physical AI bet is real and well funded. But hardware and compute travel freely; the input that stays put is EU-jurisdiction demonstration data.
6 min read
policy
Sovereign robot data: why EU storage is not enough
Storing a robot dataset on EU servers is data residency, not data sovereignty. Why the US CLOUD Act reaches EU-hosted data, and what real control needs.
7 min read
trends
Why world models entered the humanoid race
World models let humanoids learn dynamics from raw video, no robot in the loop. Why every lab is building one, and where learning from video breaks down.
6 min read
policy
Egocentric data in Europe: capture is now a spec
Why robotics converged on egocentric data, and why capturing it in Europe turns consent, provenance, and GDPR into a dataset spec buyers now ask for.
7 min read
trends
Teaching robots from internet video: promise and limits
The internet holds billions of hours of people doing things. What web video genuinely teaches a robot, the wall it hits, and the data still to be captured.
7 min read
policy
Provenance is now a spec: the EU AI Act and robot data
Under the EU AI Act, provenance is an engineering spec you build into the capture pipeline, not paperwork bolted on later. Capture-time logging wins.
7 min read
trends
The Data Flywheel in Robotics: Why the Loop Compounds
Deploy, collect, retrain: the robot data loop is real, but it only compounds when you mine failures, not hours. A field-level look at what turns the wheel.
6 min read
policy
Why capture robot training data in Europe
Capturing robot training data in Europe decides who controls the dataset, what you may legally do with it, and whether it ships under EU law.
6 min read
trends
Scaling Laws for Robotics: Does More Data Win?
Does more data mean better robot policies? Where the LLM scaling-law analogy holds for VLA models, and the three joints where it breaks down.
6 min read
trends
The European Physical AI landscape, mapped by data
A field map of European physical AI, sorted not by hardware but by where each player, from NEURA to factory-floor pilots, actually gets its training data.
6 min read
policy
Consent by design: GDPR for robot demonstration data
Egocentric robot demonstration data records a real person, so GDPR applies in full: consent, minimization, and erasure become capture-time engineering.
7 min read
trends
The Humanoid Robot Landscape in 2026, Mapped by Data
A field guide to the humanoid robot companies of 2026, mapped not by hardware but by data strategy: teleoperation, simulation, human video, shared corpora.
6 min read
trends
Where humanoids actually work on factory floors in 2026
Humanoids are on factory floors in 2026, but as narrow, supervised pilots. Where they really work, what they do, and why each pilot is also a data program.
6 min read
policy
Who owns robot data? The EU Data Act, explained
A robot, a capture rig, and a sensor array all emit machine-generated data. How the EU Data Act reshapes who can access, share, and license it.
7 min read
trends
The Humanoid Data Company: Why Proprietary Data Wins
Hardware is converging, so the humanoid moat moved to proprietary robot data. Why first-person demonstration data is the asset that compounds.
6 min read
policy
C2PA and content provenance for AI training data
C2PA was built to label AI-generated media. The same signed, tamper-evident manifest maps onto robot training data, and EU AI Act buyers will ask for it.
7 min read
trends
The data bottleneck in embodied AI
Compute and model architectures are ready. Real-world robot data is not. Inside the data bottleneck holding embodied AI back, and the pyramid behind it.
5 min read
trends
Why 2026 is the data year for humanoids
Hardware is shipping and compute is for sale. In 2026 the binding constraint on humanoid robots is training data. Why the wall moved to the data layer.
5 min read
comparisons
Synthetic Robot Data vs Real: The Mix That Works
Synthetic data now outnumbers real robot trajectories by orders of magnitude, yet the systems that generalize still blend both. How to set the ratio.
6 min read
comparisons
Open X-Embodiment vs DROID vs RH20T: How They Differ
Open X-Embodiment, DROID, and RH20T compared on scale, sensors, and provenance, and why raw trajectory counts mislead more than they reveal.
6 min read
policy
Privacy by design in human-demonstration capture
Egocentric demonstration data records a real face, home, and voice, so it is personal data by construction. Treat privacy as a capture spec, not policy.
7 min read
comparisons
Open-Source Robot Models vs Closed Foundation Stacks
A field guide to the open-source robot model debate: how the LeRobot ecosystem compares with proprietary stacks like Gemini Robotics, and who benefits.
6 min read
comparisons
Teleoperation vs human video: two roads to robot data
Teleoperation gives perfectly aligned robot demonstrations at brutal cost; human video gives cheap scale you have to retarget. The tradeoffs, side by side.
6 min read
trends
Why the robot data flywheel needs failures, not just wins
Success-only demonstrations flatter a robot policy and hide where it breaks. Why the data flywheel only compounds when you capture failures and near misses.
7 min read
comparisons
GR00T vs pi0 vs LeRobot: three data strategies
NVIDIA GR00T, Physical Intelligence pi0 and Hugging Face LeRobot each solve the robot-data shortage a different way. An architecture and data comparison.
5 min read
policy
Why human demonstration data scales better than scripted data
Scripted trajectories are clean and cheap, but coverage is what generalizes. Why human demonstration data is the more scalable base for robot policies.
6 min read
deep-dives
Mobile manipulation: data for robots that move and grab
Most robot data assumes an arm bolted to a table. Robots that walk to the object and then grab it need data that couples navigation with manipulation.
7 min read
deep-dives
Tactile Sensing: The Missing Modality in Robot Data
RGB video goes blind the instant a gripper touches an object. Here is what tactile sensing and force data add to manipulation, and why they stay scarce.
6 min read
deep-dives
Teleoperation 101: How Robots Learn From Human Hands
Teleoperation is how most robot manipulation data is made. Inside the pipeline: rigs, latency, annotation, and the QA that turns recordings into signal.
7 min read
deep-dives
Multimodal Robot Data: Force, Audio, Depth Beyond RGB
Force, contact audio, and depth carry what contact-rich manipulation runs on, the signals RGB never records. What to capture, and why timing is hard.
7 min read
deep-dives
Why dexterous hands need richer manipulation data
A two-finger gripper is a solved data problem. A five-finger dexterous hand doing in-hand manipulation is not, and the data barely exists.
7 min read
deep-dives
LeRobot: an open-source on-ramp to robot learning
How Hugging Face's LeRobot lowered the barrier to robot learning: a shared data format, sub-$200 arms, and fine-tunable policies like ACT and pi0.
6 min read
deep-dives
Hz vs fps: Robot Capture Rate and Control Frequency
Why a robot rig samples at 240 Hz but a policy acts at 30 Hz, how capture rate and control frequency differ, and what multi-rate sensing costs.
6 min read
deep-dives
How teleoperation data is collected, end to end
A full walkthrough of the teleoperation data pipeline, from leader-follower rig to a synchronized, filtered, labeled trajectory, and where QA is won.
6 min read
comparisons
How to actually measure a robot policy
A robot policy can post 90% on a benchmark and still fail in the field. The pitfalls of policy evaluation, and how to test one honestly.
7 min read
deep-dives
How Much Data a Humanoid Skill Needs: Sample Efficiency
How many demonstrations does a humanoid robot skill really need? An honest range across imitation learning, pretrained policies, and human video.
6 min read
policy
The EU AI Act and Robot Training Data: What to Know
What the EU AI Act actually requires of embodied training data: Article 10 quality rules, GPAI provenance duties, and the 2026-2027 timeline for robots.
7 min read
foundations
Stretching scarce demos: data augmentation for robot policies
Real robot demos are scarce, so teams stretch them with data augmentation. What crops, diffusion, and replay add for free, and what they cannot fake.
7 min read
policy
The Economics of Robot Data: Cost per Demonstration Hour
Teleoperation, simulation, and egocentric video each carry a different cost per demonstration hour. Here is why that number drives robot-data strategy.
6 min read
deep-dives
From demonstration to deployment: the robot training pipeline
Pretrain, post-train, sim2real, and evaluation are one loop closed by data and honest testing, not a checklist. A field guide to the robot training pipeline.
6 min read
deep-dives
Data Quality vs Quantity in Robot Learning
Why coverage beats raw hours in robot learning: TRI's diversity findings, how to read a dataset for data quality, and the real tradeoffs.
6 min read
policy
Active learning: collect the right robot data, not just more
Robot demonstrations are expensive, so collect the ones a policy learns most from: active learning by uncertainty, disagreement, and deployment failures.
7 min read
policy
Why Data Jurisdiction Belongs in Your AI Training Pipeline
The EU AI Act makes training-data provenance an engineering problem, not a legal one, for robot-learning datasets that carry jurisdiction and consent.
6 min read
deep-dives
What contact-rich manipulation demands from data
Force and tactile signals, not pixels, decide contact-rich manipulation. Why RGB can never supply them, and why touch data is the real bottleneck.
6 min read
policy
Why egocentric human video is among the cheapest robot data
Egocentric video takes the robot out of the capture loop, so the hours come cheap. Why it scales, and the embodiment-gap catch that keeps it honest.
7 min read
deep-dives
Retargeting: mapping a human demonstration onto a robot body
A human hand has twenty-plus joints; a robot gripper has two. Retargeting rewrites a demonstration into the robot's body, and this is where it leaks.
7 min read
deep-dives
The hidden cost of robot data annotation
Language labels, calibration, and time-sync are the real cost of robot data annotation. Why headline trajectory counts hide whether a dataset is trustworthy.
6 min read
foundations
What is egocentric data, and why robots learn from it
First-person capture from head and wrist cameras teaches robots to grip better than any wall camera. What egocentric data is, and the catch.
6 min read
deep-dives
Running a VLA in real time: the inference latency problem
A control policy must act faster than the world changes, yet a big VLA needs 100+ ms per forward pass. The techniques that close the gap in real time.
7 min read
foundations
What is a robot foundation model? VLAs explained
GR00T, pi0 and Gemini Robotics are robot foundation models, not chatbots. What a VLA is, how it differs from an LLM, and why data is the hard part.
5 min read
foundations
Vision-Language-Action Models: A Plain-English Primer
How a vision-language-action model turns camera pixels and a spoken instruction into robot motion, without the math: the models, datasets, and real limits.
6 min read
foundations
The sim-to-real gap, explained: where real data wins
Simulation trains robots cheaply, but it quietly underfits contact, deformables, and human context. A field guide to where real data is non-negotiable.
6 min read
foundations
Post-training a VLA: from generalist to specialist policy
Fine-tuning turns a generalist robot policy into one you can deploy. A concrete look at VLA post-training: methods, data, failure modes, and evaluation.
6 min read
foundations
Whole-body control: how humanoids coordinate legs and arms
A humanoid must balance on two legs while it manipulates, so every reach shifts its own center of mass. Why whole-body control is hard, and data-starved.
6 min read
foundations
Imitation vs Reinforcement Learning for Manipulation
Why most robot manipulation stacks learn by imitation first and reach for reinforcement learning second, when each wins, and how the two combine.
6 min read
foundations
Gemini Robotics, explained: embodied reasoning in a VLA
DeepMind split Gemini Robotics into a reasoning brain and an acting body. How embodied reasoning turns a VLA from a memorizer into a planner.
6 min read
deep-dives
The tactile sensors giving robot hands a sense of touch
Capacitive arrays, barometric MEMS, and GelSight-style gels each give robot hands touch a different way. A field guide to tactile sensor hardware.
7 min read
foundations
Diffusion policy, explained: generative robot control
Why generative models, diffusion and flow matching, became the default way robot foundation models produce multimodal action, and what it asks of the data.
6 min read
foundations
Cross-Embodiment Learning: One Policy, Many Robots
How a single policy learns to drive many different robots, why their heterogeneous bodies make it hard, and what Open X-Embodiment actually proved.
7 min read
foundations
Action tokenization in a VLA, explained
How vision-language-action models turn continuous motion into discrete action tokens a transformer can predict, and why that choice caps resolution.
6 min read