robot-atlas
An encyclopedic, interactive guide to modern robotics for ML engineers, from learned manipulation policies to the classical stack underneath them.
The six core domains
Manipulation & Learned Policies
From behavior cloning to vision-language-action models: how modern robots learn to act.
12 of 12 modules published
RL, Sim-to-Real & Locomotion
Reinforcement learning at scale, massively parallel simulation, and the transfer problem.
6 of 6 modules published
World Models
Learned simulators, action-conditioned video prediction, and the JEPA counterargument.
5 of 5 modules published
Data, Hardware & Evaluation
The embodied data bottleneck, the machines themselves, and the measurement crisis.
5 of 5 modules published
Classical Foundations
Kinematics, planning, control, estimation, and grasping: the stack under every learned policy.
5 modules planned
Frontier & Open Problems
Reliability, dexterity, generalization, and the competing theses about what wins.
5 modules planned
Adjacent domains
Adjacent Domains
4 modules plannedAutonomous vehicles, drones, surgical robotics, and space robotics in brief.
- Autonomous Vehicles
- Drones and Aerial Robotics
- Surgical Robotics
- Space Robotics
Interactive tools
Market Map
The embodied-AI industry as data: more than a hundred companies across foundation models, humanoids, industrial systems, vertical applications, simulation, and components, filterable by approach, geography, stage, and funding.
6 segments, 100+ companies
3D Kinematics Playground
Move a real robot arm: joint-slider forward kinematics, click-to-reach inverse kinematics, and trajectory replay.
FK, IK, replay
How to read this atlas
Six core domains form the spine: Manipulation & Learned Policies, reinforcement learning and sim-to-real, world models, data, hardware and evaluation, classical foundations, and the frontier of open problems. A seventh group, Adjacent Domains, sketches vehicles, drones, surgical, and space robotics in brief.
Modules stand alone, but within a domain they build on each other in registry order. If you come from ML, start with the first published module, Action Chunking (ACT and ALOHA): precise prose, inline citations to primary sources, and a live interactive you can manipulate. That is the format every module follows. Every non-obvious claim carries a citation chip that links to the paper, lab writeup, or official documentation behind it.
Planned modules appear in the taxonomy before they are written, so the sidebar doubles as the roadmap. Draft entries are marked planned and go live as they are reviewed.
Featured interactive
A 95% per-step success rate sounds strong. Compounded over a 30-step episode it is not. Move the sliders to see how small per-step errors erode end-to-end reliability; the Frontier domain develops the argument.
(0.950)^30 = 21.5% episode success