The reading room
Robotics' End Game
The big picture — where robotics and embodied AI are headed — from NVIDIA's GEAR lead. Start here.
Inside the World's Smartest Robot Brain
A beautifully made primer on how a vision-language-action model turns camera pixels into motor commands.
Humanity's Last Machine
A gorgeous, component-level tour of what actually goes into a humanoid robot.
The Hugging Face Robotics Course
Free, hands-on, built by the LeRobot team. The gentlest on-ramp — begin at Unit 0.
Robot Learning: A Tutorial
One page that maps the whole field — data, imitation, policies — in an afternoon.
LeRobot — the library
The open engine our whole stack runs on. Motors, cameras, datasets, training, inference.
Training ACT on the SO-101
A candid, failure-and-all end-to-end walkthrough: setup, teleop, data, and training ACT on our exact arm.
RL to grasp and lift, in MuJoCo
Trains SAC to lift a cube in simulation. The eleven versions of the reward table are the real lesson.
Grasping from pixels
The same task from camera images instead of privileged state. The step where it starts to resemble what we run.
HIL-SERL: real-world grasping from scratch
On hardware, no simulator, a human correcting it as it learns. The HIL-SERL machinery that ships inside our LeRobot install.
Fine-tuning SmolVLA
Pick-and-place with a vision-language-action model on our arm. Where three months of one person's work lands.
MuJoCo Menagerie — trs_so_arm100
The official MJCF model of our arm. It ships two calibration variants: pin one, or sim and hardware silently disagree.
lerobot-sim2real
Train with RL in GPU-parallel ManiSkill3, then deploy to a real arm zero-shot. Built on LeRobot, so hardware is solved.
Lil'Log
The clearest long-form deep-dives in ML — diffusion, RL, agents, and more.
Yang Song
Score-based diffusion and consistency models, explained by the researcher who introduced both. The clearest writing on why diffusion works.
Sander Dieleman
Deep, beautifully written essays on diffusion and generative models, usually the best explanation of an idea anywhere.
SIGReg, derived from scratch
A rigorous, beautifully built derivation of LeJEPA's anti-collapse regularizer — characteristic functions to a trainable loss.
Perry Dong
RL for real robots, from the author of EXPO. The post-training argument, and a tuning guide for it.
Behavioral cloning mystery
Overfitting helps, open-loop beats closed-loop, pick-and-place wants a 0.5B MLP. Four things cloning should not do, all from distribution shift.
Diffusion Policy, explained by its author
The method, the intuition, and the honest open questions — straight from the source.
ALOHA & ACT — a walkthrough
Alexander Soare breaks down action chunking and the ALOHA setup, end to end.
VQ-BeT, explained by its authors
Mahi Shafiullah & Seungjae Lee tokenize actions with a residual VQ-VAE — as expressive as diffusion, ~5× faster.
OpenVLA — a walkthrough
Moo Jin Kim, OpenVLA's author, on what it takes to build and release an open vision-language-action model.
Embodied Intelligence — the cost of a new behaviour
Teleop versus sim-to-real, distillation, scanning rooms into simulators, adversarial safety testing. His unit throughout is human hours, not success rate.
Question the role of vision for robotic control
Vision is for planning; control needs force. Why manipulation lags locomotion, and why a hand heavier than its payload fails.
Train an ACT policy on the SO-101
A practical, start-to-finish walkthrough of training ACT on the SO-101 with LeRobot.
Ilia Larchenko
Deep, practical robotics — SO-101 teleop, training, and reinforcement learning on the arm.
Dan Makes Things
An engineer who actually builds things, SO-101 included. Watch a whole build come together.
Foundation Robotics
Short, sharp explainers on frontier hardware — robotic hands, humanoids, actuators. Good for five spare minutes.
The Robot Brains
Long, generous interviews with the researchers and founders building robot brains. Running since 2021, and still the deepest of them.
Automated
Weekly conversations with robotics and humanoid founders, from TechCrunch's longtime robotics editor. Closer to the business than the research.
Over the Horizon
Technical, opinionated deep-dives on where the humanoid industry is going, usually with the founders doing the building.
Embodied AI 101
A new episode most days, each one walking through a single recent paper on manipulation, teleoperation or embodied models.
Essence of Linear Algebra
The visual foundation — vectors, matrices, and what they actually mean. Before any ML.
Generative Diffusion Models
A hand-picked reading order through diffusion, flow matching and optimal transport. The maths under Diffusion Policy and the VLAs.
Modern Robotics
Mechanics, planning and control: the classical robotics foundation, taught clearly and paired with a free textbook and code library.
Robotic Manipulation
Effectively the textbook for what we do. Free notes, runnable Drake notebooks, chapters from pick-and-place to tactile sensing. Start here.
Underactuated Robotics
Dynamics and control for real, underactuated robots. Deep, rigorous, and worth it.
ECE 4560 — Intro to Robotics
Ten labs with a MuJoCo repo, taught on SO-101 arms. Kinematics, planning, Lie groups: a ready-made curriculum for our hardware.
CS 285 — Deep RL
The canonical graduate deep reinforcement learning course, from policy gradients to offline RL. Every lecture and homework is free online.
6.8200 — Computational Sensorimotor Learning
RL, imitation, self-supervised and model-based learning for control. Hands-on first half, open research questions second.
CIS 6280 — World Models
Twenty-three lectures from VAEs to video generation to control. Slides are public.
CS 224R — Deep RL
Deep RL with a robot-learning bent, from one of the field's leading researchers.
Read something great we missed?
This list is curated, not complete. Members keep it sharp — send a link and a one-line why.