๐ Introducing SARM: Stage-Aware Reward Modeling for Long-Horizon Robot Manipulation
Robots struggle with tasks like folding a crumpled T-shirtโlong, contact-rich, and hard to label. We propose a scalable reward modeling framework to fix that. 1/n
Start learning LLM papers and all the pre-training post-training tricks, afraid of that one day all robotics researchers can do is data cleaning and tuning rollouts ๐คฃ๐คฃ
Introducing GEN-0, our latest 10B+ foundation model for robots
โฑ๏ธ built on Harmonic Reasoning, new architecture that can think & act seamlessly
๐ strong scaling laws: more pretraining & model size = better
๐ unprecedented corpus of 270,000+ hrs of dexterous data
Read more ๐
@JasonMa2020's previous reward model works (VIP, LIV etc.) as well as @DynaRobotics 24-hour folding towel blog were definately inspirations to SARM. PLease checkout Dyna's new folding T-shirt demo at #CoRL2025x.com/JasonMa2020/stโฆ
6/n
We just did Worldโs first on-stage autonomous demo of long-horizon dexterous VLA ๐จ
No training. No setup. Performance out of the box. Live demo is hard and unpredictable, but we felt great about our modelโs generalization, and it went pretty well!
๐ฏ Zero-shot. 100% success.
โYou canโt make progress until you are able to measure it. Robotics still doesnโt have such a rallying call. No one agrees on anything.โ I ๐ฏ agree with the recent post from @DrJimFan.
To break this impasse, we are excited to announce ManipulationNet (manipulation-net.org), a
Incredible results! I donโt feel like I can doing those tasks for 2 hours without failure. Real world RL, though hard, will be the path leading to super human reliability VLA!
Introducing RL-100: Performant Robotic Manipulation with Real-World Reinforcement Learning. lei-kun.github.io/RL-100/
7 real robot tasks, 900/900 successes. Up to 250 consecutive trials in one task, running 2 hours nonstop without failure.
High success rate against physical
๐ง SARM = Stage-aware reward model
Instead of labeling every frame manually, we leverage language annotations to learn stage & progress.
โ Robust to noisy demos
โ Generalizes to policy rollouts
โ Enables semantically meaningful progress
2/n
Today, we present a step-change in robotic AI @sundayrobotics.
Introducing ACT-1: A frontier robot foundation model trained on zero robot data.
- Ultra long-horizon tasks
- Zero-shot generalization
- Advanced dexterity
๐งต->
๐ Robust to OOD + noisy policy rollouts
SARM isnโt just for clean human demos. It accurately estimates progress even on noisy, OOD policy rollouts.
๐ a demo where SARM detects recession and not cheated by โfake finishโ. 3/n
Though I canโt attend #IROS2025 due to visa issues, Iโm thrilled two of my papers will be presented!
1๏ธโฃ GRaD-Nav: drone visuo-motor RL with 3DGS (Wed Oct 22, 10:30, Rm 311A)
2๏ธโฃ GRFM-Net: autotuning bipedal MPC (Thu Oct 23, 16:40, Rm 101)
#RobotLearning#ReinforcementLearning
โ๏ธ๐ค What if an embodiment-agnostic visuomotor policy could adapt to diverse robot embodiments at inference with no fine-tuning?
Introducing UMI-on-Air, a framework that brings embodiment-aware guidance to diffusion policies for precise, contact-rich aerial manipulation.
Our model can now learn from its own experience with RL! Our new ฯ*0.6 model can more than double throughput over a base model trained without RL, and can perform real-world tasks: making espresso drinks, folding diverse laundry, and assembling boxes.
More in the thread below.
