Robot Wiki

Citation-first encyclopedia of modern robot learning.

Robot Wiki is an encyclopedia of modern robotics for engineers who already know machine learning: learned manipulation policies, sim-to-real reinforcement learning, world models, teleoperation data pipelines, and the classical control stack underneath. Technical claims should be traceable to cited evidence. Sources range from research papers to first-party documentation and labelled community estimates, and a citation is not a guarantee of verification. The centre of gravity is robot learning, not a catalogue of the industry.

Start reading How to read this wiki

Domain index

Manipulation & Learned Policies
From behavior cloning to vision-language-action models: how modern robots learn to act.
RL, Sim-to-Real & Locomotion
Reinforcement learning at scale, from massively parallel simulation to the transfer problem.
World Models
Learned simulators and action-conditioned video prediction, with the JEPA counterargument.
Data, Hardware & Evaluation
The embodied data bottleneck and the machines themselves, with the measurement crisis behind every benchmark claim.
Classical Foundations
The kinematics, planning, control and estimation stack that sits under every learned policy.
Frontier & Open Problems
The reliability gap that keeps demos out of production, and the competing theses about closing it.
Adjacent Domains
Autonomous Vehicles, Drones and Aerial Robotics, Surgical Robotics, and Space Robotics.

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 essays develop the argument.

Per-step success95.0%

Episode length30 steps

(0.950)^30 = 21.5% episode success

At 95.0 percent per-step success, episode success is 21.5% at 30 steps and 0.6% at the 100-step end of the plotted range, crossing 50 percent at 14 steps as the per-step odds compound over the episode length.

Sampled episode success by episode length

episode length	episode success
0 steps	100.0%
10 steps	59.9%
25 steps	27.7%
50 steps	7.7%
75 steps	2.1%
100 steps	0.6%

Real hardware

The policies and controllers this wiki covers run on physical machines. Every photograph and diagram on the site is licensed and credited, and the full list lives on the credits page.

A yellow Boston Dynamics Spot quadruped robot walks across an airfield tarmac alongside Royal Air Force airmen in camouflage uniform. — A Boston Dynamics Spot quadruped walks with Royal Air Force airmen at RAF Leeming during the Agile Liberty 21-2 technology trials in August 2021.

Interactive tools

shoulder_pan220°
shoulder_lift200°
elbow_flex193.7°
wrist_flex190°
wrist_roll320°
gripper110°

The shipped so101.urdf defines 6 revolute joints from base_link to the moving jaw, and at the zero configuration the chain reaches 407 mm from the base with the wrist 234 mm above it; joint travel spans 320 degrees at wrist_roll and 110 degrees at gripper.

Joint travel declared by the shipped URDF

joint	lower	upper
shoulder_pan	-110°	110°
shoulder_lift	-100°	100°
elbow_flex	-96.8°	96.8°
wrist_flex	-95°	95°
wrist_roll	-157.2°	162.8°
gripper	-10°	100°

3D Kinematics Playground

Drive this arm in the browser: joint-slider forward kinematics, click-to-reach inverse kinematics, and trajectory replay.

Market Map

The embodied-AI industry as data: 111 companies across 6 segments, filterable by approach, geography, stage, and funding.

Learning paths

Pick a sequence when you want a coherent route through the wiki. Each path starts with the concepts later articles assume.

Robot learning and VLA

From embodied data and imitation learning to action models, evaluation, and deployment.

Classical robotics foundations

The dependency chain underneath learned policies: geometry, sensing, estimation, planning, and control.

World models for robotics

Start with the overloaded term, then compare latent, representation-space, video, and simulation approaches.

How to read this wiki

Modules stand alone, but inside a domain they build on each other in registry order: later entries assume the earlier ones. If you come from machine learning rather than robotics, start with Action Chunking (ACT and ALOHA). It shows the format every module follows: precise prose, inline citations, and a live interactive you can manipulate.

The prerequisites are fluency in machine learning, not a robotics background. When a module needs a classical result it says so and links to the entry that derives it, so you can read forward and backfill as needed. Terms of art are defined where they first appear and collected in the glossary.

Citation chips link claims to their named sources, including explicitly labelled community estimates where used, and the full bibliography sits at the end of each module. Where serious researchers disagree, the text names who holds which position.