Still thinking about the CoRL debate—I was on the “no” side. With ATLAS, we connect a VLM to a whole-body controller and let the robot learn from its own attempts, without training a VLA in between.
What we showed at CoRL 2026: a humanoid with no System 1. An agent-written harness linking System 2 and System 0, plus one human video demonstration, is enough to get the task done.
https://t.co/C9ZsIyBV0M
I have witnessed how much hard work and heart went into SuperMimic by Qiayuan this year. So honored to be his friend and collaborator, and so well deserved!
Reward is always a bottleneck for robotic policy finetuning. Dense rewards introduce bias or require laborious annotations, while still struggling with behavioral nuances. How to achieve dense rewards with minimal annotations at test time? https://t.co/uPIzdGdj6J #CoRL2026
Wanted to get into manipulation. Decided to first build a full stack that scales.
Plan: copy @GeneralistAI's gripper. Fun weekend project.
It was not.
So @t_k_233 and I documented every design choice and lesson along the way.
Meet Specialist: fully open-source hardware, data pipeline, training, and deployment code.
https://t.co/aoN5x0Jp6I
Can you believe these are open-loop rollouts?
Excited to introduce CARD 🃏 — Contact-Anchored Retargeting for Dexterous Manipulation.
Almost every method today is limited to kinematics alone as the 𝐢𝐧𝐭𝐞𝐫𝐟𝐚𝐜𝐞 between the human and the robot. With CARD, we track real world contact and forces alongside kinematics to perform unprecedented dexterous tasks through tactile reactive replay!
CARD is scalable:
📹 Multiple input sources
🤖 Multiple robot hands
⚡ needs no object models, no simulator, 10 Hz
Fine-tuning a general world model in the real world seems substantially harder. FADA adapts an inverse-dynamics map from observed trajectories, so it does not require estimating the full hybrid state or identifying contact modes. A general humanoid world model would, and that state-estimation problem is already extremely challenging in contact-rich systems—likely also one reason DIAL-MPC has not yet shown real-world humanoid results. On the other hand, learned world models naturally smooth hybrid dynamics unless discontinuities are modeled explicitly, which give APG better-conditioned Jacobians than an exact differentiable simulator.
It is a new paradigm -- world model + APG. This enables 1000x faster training of generalist whole-body tracker. And opus 5.5 high can reproduce it within 1h for lafan.
Recently I wonder if we are turning dumb and falling into scaling slop. We need more @qiayuanliao .
I never really worked on motion tracking before because I saw it as largely demo-driven research.
SuperMimic changed how I think about it. Scaling to 129K+ motions turns motion tracking from “can we imitate this clip?” into a much more fundamental question: 𝗵𝗼𝘄 𝗱𝗼 𝘄𝗲 𝗯𝘂𝗶𝗹𝗱 𝘀𝗰𝗮𝗹𝗮𝗯𝗹𝗲, 𝗴𝗲𝗻𝗲𝗿𝗮𝗹-𝗽𝘂𝗿𝗽𝗼𝘀𝗲 𝗵𝘂𝗺𝗮𝗻𝗼𝗶𝗱 𝗰𝗼𝗻𝘁𝗿𝗼𝗹𝗹𝗲𝗿𝘀 𝘄𝗶𝘁𝗵 𝗿𝗲𝗮𝘀𝗼𝗻𝗮𝗯𝗹𝗲 𝗰𝗼𝗺𝗽𝘂𝘁𝗲?
We found a way to scale humanoid control that reduces GPU-hours by more than 1000×.
Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.
Project: https://t.co/6gc9WKm9qY
Demo: https://t.co/OHRvx5ogQY
Can you sample a robot policy the way you sample an image? 🎲
No RL run. No reward. No fine-tuning.
We train a diffusion model on the 𝘄𝗲𝗶𝗴𝗵𝘁𝘀 of thousands of RL policy networks, then sample brand-new ones that walk 🚶, dance 🕺, and spin-kick 🦵 on a real Unitree G1 🤖
The best part? The 𝘄𝗲𝗶𝗴𝗵𝘁𝘀 have a secret 👀 Every network, no matter the motion, pays the most attention to the hips and waist — embodiment comes first; motion comes second. 🦾
📄 arXiv: https://t.co/iC8g2eAzYB
🌐 Website: https://t.co/xH5bdm9TZs
#Robotics #HumanoidRobots #DiffusionModels #RobotLearning #EmbodiedAI
Practice makes perfect. A truly agentic robot shouldn’t just know what to do. It should also figure out how to get better.
What if a robot could watch a demonstration, figure out what it needs to practice, and learn how to get better through practice?
Introducing RPG: Guided Self-Improvement for Embodied Agents, a real-to-sim-to-real pipeline that turns real-world demonstrations into targeted practice.
RPG reconstructs practice tasks from demonstrations, uses demo videos to guide skill improvement, and continually updates a shared skill system from failures.
On real robots, RPG achieves 30/30 successes across 3 tasks. In simulation, it reaches 95.0% success across 22 manipulation tasks after 15 rounds of self-improvement.
Demo → Reconstruct → Guided Practice → Self-Improve → Go Real
Can a humanoid learn robust whole-body loco-manipulation directly from a human? No robot, no teleop, no retargeting.
Meet Workhorse 🐴: our G1 topples and climbs a suitcase, catches a thrown box, and sorts boxes with hands and a kick. Fully autonomous.
More videos at https://t.co/HrxCv199F7
@xiaolonw@chris_j_paxton Any online optimization (IK, ID, MPC) for low-level control should be abandoned. A RL policy can fundamentally track more precisely, more robustly, and adaptively to unknown disturbances and uncharacterized noise, especially for low-cost hardware, even on a manipulator.
sharing our new project Smash!🏓
We built the first outdoor humanoid table tennis player with fully onboard perception — no MoCap, no external cameras. check the video for details.
paper and code on the way!
Happy to see my code could solve the problems my labmate previously refused to tackle until I made it easy.
Funny how it turned into a scoop of my own PhD thesis topic, but glad the tools work. 🙃