PhD candidate at ETH RSL @leggedrobotics.
Joined March 2025
- Tweets294
- Following308
- Followers1.7K
- Likes305
Pinned Tweet
Introducing Humanoid Mobility via Attention-based Contacts!
One policy enables diverse motions over challenging terrains without reference motions at deployment.
No manually designed transition motions
No skill switching mechanism
No manual data augmentation
13.2 minutes of data
Click a goal and go!
paper: tbd
website: leggedrobotics.github.io/ACP…
Junzhe (JJ) He retweeted
Vision foundation models like DINO and SigLip now power much of robot learning.
But they are built for RGB. What about depth?
Introducing DeFM: a foundation model for depth images, built for robotics.
🌐 de-fm.github.io
Weights, code, and demo are out today 🧵
Very impressive!
Made a demo for MotionBricks running 100% in your browser.
You steer a robot, change its walking style or approach an obstacle. The app turns that into target poses, and the model uses its recent poses to generate the motion connecting them.
It outputs 3D joint motion, runs locally on your CPU.
huggingface.co/spaces/townbo…
Who vaults better? Me or G1?
Introducing Humanoid Mobility via Attention-based Contacts!
One policy enables diverse motions over challenging terrains without reference motions at deployment.
No manually designed transition motions
No skill switching mechanism
No manual data augmentation
13.2 minutes of data
Click a goal and go!
paper: tbd
website: leggedrobotics.github.io/ACP…
Junzhe (JJ) He retweeted
With @JayHe748646 @KehanWen170077 @leggedrobotics , we push humanoid mobility to next frontier.
Achieved by a generalist contact planner + a generalist tracker + a generalized neural mapping, warmstarted from few human data.
Website: leggedrobotics.github.io/ACP…
Promo Video 👇
We demonstrate our approach on the Unitree G1 humanoid robot through extensive real-world experiments, including ladder climbing, vaulting, crawling, sparse-terrain traversal, seamless skill transitions, and autonomous failure recovery. Quantitative evaluations further demonstrate strong generalization to unseen and composite terrains, achieving a 92.5% average success rate on composite courses compared with 58.8% for a trajectory-generation-based baseline.
Our key contribution is an attention-based contact planner that predicts future body–environment interactions, coupled with a contact-conditioned whole-body controller. By using motion data solely as a training prior and subsequently fine-tuning through reinforcement learning, our framework enables a single policy to autonomously adapt to unseen terrains without requiring reference motions at deployment.
Achieving versatile humanoid mobility in complex environments remains a fundamental challenge in robotics. While reinforcement learning offers strong adaptability, discovering coordinated whole-body behaviors remains difficult. Conversely, motion imitation provides natural coordination but often limits generalization beyond reference trajectories. With @ChongZzZhang and @KehanWen170077, we address this trade-off through a contact-centric learning framework that combines the strengths of both approaches.
Junzhe (JJ) He retweeted
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: supermimic.github.io
Demo: supermimic.github.io/demo/
Junzhe (JJ) He retweeted
one robot policy, trained from scratch, solves all 20 NIST assembly tasks. no demos 🤯
AlphaNIST (NVIDIA + UW) only gets told one thing: assembled or not.
🎯 98.2% success across all 20 tasks in sim, no per-task tuning
🔩 the 3 tiniest parts failed when trained alone. training on all 20 at once cracked them
🧠 chains 10 assemblies in a row, rewarded once at the very end
Why it matters: RL may find manipulation strategies we'd never script.
Oh wow this is actually really cool!! Although I believe ame2 can also solve this type of terrains without expert policies distillation, achieving this by rather a simple modification on ppo without massive reward tuning is quite impressive.
Our manipulation results were mind-blowing the first time I saw them, but earlier on, my favorite result was this locomotion run. Many great people have done parkour-type skills with locomotion already, but as far as I am aware, most, if not all, of these works trained separate policies for each terrain, and then distilled into a single policy. They also typically contained tons of hand-engineered rewards and curriculums.
A very early motivation for SGS was to simply figure out if we could get similar results but get rid of as much of this hand-engineering as possible, and we managed to achieve that! This video shows a continuous 1-min run in much harder terrains than ever before, all solved by the same policy.
Junzhe (JJ) He retweeted
New work led by @shinben0327 on HOI-Retarget!
We retarget human–object interactions to different humanoid morphologies by preserving object contacts, rather than only matching the human pose. Really fun project, we’re releasing both code and dataset.
Presenting HOI-Retarget: Contact-Centric Retargeting for Human–Object Interaction.
🌐 Website: shinben0327.github.io/hoi-re…
📄 Paper: arxiv.org/pdf/2609.34674
👥 Authors: @shinben0327, @alopeze99, @JayHe748646, @MatthiasCaro, Marco Hutter
@leggedrobotics @ETH_en
Junzhe (JJ) He retweeted
Presenting HOI-Retarget: Contact-Centric Retargeting for Human–Object Interaction.
🌐 Website: shinben0327.github.io/hoi-re…
📄 Paper: arxiv.org/pdf/2609.34674
👥 Authors: @shinben0327, @alopeze99, @JayHe748646, @MatthiasCaro, Marco Hutter
@leggedrobotics @ETH_en
Junzhe (JJ) He retweeted
Incredibly excited to return to ETH Zurich as Assistant Professor of Robotics and Artificial Intelligence next summer. I'll be recruiting PhD students to build a new lab on general-purpose embodied intelligence for humanoid robots and dexterous manipulation. Details coming soon!
