Gu Zhang (@Gu__Zhang) header
Gu Zhang (@Gu__Zhang) profile picture

CS PhD @Tsinghua_IIIS | Prev. Student Researcher @MIT_CSAIL @SJTU1896 | Research interest focus on dexterous manipulation

Joined August 2021
Photo by Gu Zhang (@Gu__Zhang)Photo by Gu Zhang (@Gu__Zhang)Photo by Gu Zhang (@Gu__Zhang)Photo by Gu Zhang (@Gu__Zhang)Photo by Gu Zhang (@Gu__Zhang)Photo by Gu Zhang (@Gu__Zhang)Photo by Gu Zhang (@Gu__Zhang)Photo by Gu Zhang (@Gu__Zhang)Photo by Gu Zhang (@Gu__Zhang)Photo by Gu Zhang (@Gu__Zhang)
What if we could teach dexterous robots through natural human demonstrations—with human data only? Introducing TouchBridge: transferring contact-rich skills from human hands to robot hands through touch. 🖐️🤖 📄 arXiv: arxiv.org/abs/2610.40349 🌐 Website & demos: touchbridge2026.github.io/ Across 4 real-world tasks, TouchBridge achieves 72.5% average success using human data only, compared with 77.5% using the same number of teleoperated robot demonstrations. Human demonstration collection is 7.0× faster. The key: capture how fingers contact an object, then preserve those contacts when transferring the demonstration to a different hand. No object meshes, object-pose estimates, or physics simulation are needed for retargeting. From turning book pages to screwing in a light bulb—here’s how it works. 🧵👇
6
12
2
67
4,092
Congrats to Wenxuan's Latest work! Really impressive to see SLA pushing the boundaries of spatial-language-action models through continual pre-training. Robust real-world execution and error correction are key steps toward more capable and general-purpose robots. Congrats on the great work! 🚀
SLA: Scaling Spatial-Language-Action Model via Continual Pre-training Real-robot SLA exhibits robust performance and error-correction capabilities in continuous task execution. The system continues to function effectively on physical hardware. Both the code and model weights are fully open-sourced. We welcome everyone to try it out. Arxiv: arxiv.org/pdf/2611.08373 Homepage: embodiedsla.github.io/ Github: github.com/OpenHelix-Team/Sp… Huggingface: huggingface.co/papers/2611.0… #Robotics
1
161
Gu Zhang retweeted
Everyone is saying robotics is solved, but from my view it is actually the start, the start of general-purpose robots. Every single day and single month we are getting closer and closer to that great goal
Replying to @claudeai
Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do. On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots.
4
4
75
3,557
Congrats to @fancy_yzc latest work! Real-world improvement is essential part for reliable deployment. Very impressive work!
Reliable real-world manipulation requires policies that can improve from deployment experience. Introducing PIVOT: Policy Improvement via Trust-Region Optimization from Deployment Experience — accepted to CoRL 2026. PIVOT combines demonstrations, human corrections, and autonomous rollouts in an iterative post-training loop, producing consistent gains across precise, contact-rich bimanual tasks.
1
6
970
Gu Zhang retweeted
Touch can server as a bridge connecting human data and dexterous robots. We are excited to introduce our recent project Touchbridge. Arxiv: arxiv.org/abs/2610.40349 Project: bridge2026.github.io
What if we could teach dexterous robots through natural human demonstrations—with human data only? Introducing TouchBridge: transferring contact-rich skills from human hands to robot hands through touch. 🖐️🤖 📄 arXiv: arxiv.org/abs/2610.40349 🌐 Website & demos: touchbridge2026.github.io/ Across 4 real-world tasks, TouchBridge achieves 72.5% average success using human data only, compared with 77.5% using the same number of teleoperated robot demonstrations. Human demonstration collection is 7.0× faster. The key: capture how fingers contact an object, then preserve those contacts when transferring the demonstration to a different hand. No object meshes, object-pose estimates, or physics simulation are needed for retargeting. From turning book pages to screwing in a light bulb—here’s how it works. 🧵👇
1
2
176
Gu Zhang retweeted
Kinematic retargeting can produce similar motion, but once object interaction is involved, tiny inaccuracies get amplified—it can’t model contact. Using tactile info as an intermediary can greatly mitigate this, and it mitigates the need for RL retargeting via real2sim2real.
What if we could teach dexterous robots through natural human demonstrations—with human data only? Introducing TouchBridge: transferring contact-rich skills from human hands to robot hands through touch. 🖐️🤖 📄 arXiv: arxiv.org/abs/2610.40349 🌐 Website & demos: touchbridge2026.github.io/ Across 4 real-world tasks, TouchBridge achieves 72.5% average success using human data only, compared with 77.5% using the same number of teleoperated robot demonstrations. Human demonstration collection is 7.0× faster. The key: capture how fingers contact an object, then preserve those contacts when transferring the demonstration to a different hand. No object meshes, object-pose estimates, or physics simulation are needed for retargeting. From turning book pages to screwing in a light bulb—here’s how it works. 🧵👇
1
1
3
315
Why this matters: contact-rich dexterity has been bottlenecked by expensive robot teleoperation. TouchBridge turns fast, natural human demonstrations into robot-ready data while preserving contact, a practical path to scaling robot learning from humans.
1
1
85
Want to see the full pipeline—from human touch to autonomous robot manipulation? 📄 Read the paper: arxiv.org/abs/2610.40349 🌐 Explore the videos, results, and interactive method demos: touchbridge2026.github.io/ Thanks to the great efforts from co-leaders @HaoZhe1451 and Zirui Zou. Also thanks to the great advice and support from Yuzheng Chen, Jinlong Wu, Zixuan Hu, Jianke Zhang, Yucheng Hu, @fancy_yzc and Prof. @HarryXu12 !
3
97
Gu Zhang retweeted
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
10
53
15
229
167,254
Gu Zhang retweeted
It turns out robots are just computers. Very anticlimactic.
7
8
3
111
14,539
Most dexterous demo ever seen! Human data finally actually work! Congrats to the great team!
We are back. After one year of quiet building. Introducing GENE-26.5, our first robotic brain that takes a major step toward human-level capability. For years, robotics has struggled to learn from the world’s largest and valuable data source: Humans. Solving it means rethinking the whole stack from the ground up: - A robotics-native foundation model. - A 1:1 human-like robotic hand. - A noninvasive data collection glove for motion, force, and touch. - A simulator that turns weeks of experiments into minutes. GENE-26.5 is trained across language, vision, proprioception, tactile, and action. We designed a set of tasks to test how far we can go with this new paradigm. Fully autonomous, 1x speed, one model, same weights. (Enjoy with sound on) We are approaching the endgame for robotics. And this is just a beginning.
5
416
Gu Zhang retweeted
We are back. After one year of quiet building. Introducing GENE-26.5, our first robotic brain that takes a major step toward human-level capability. For years, robotics has struggled to learn from the world’s largest and valuable data source: Humans. Solving it means rethinking the whole stack from the ground up: - A robotics-native foundation model. - A 1:1 human-like robotic hand. - A noninvasive data collection glove for motion, force, and touch. - A simulator that turns weeks of experiments into minutes. GENE-26.5 is trained across language, vision, proprioception, tactile, and action. We designed a set of tasks to test how far we can go with this new paradigm. Fully autonomous, 1x speed, one model, same weights. (Enjoy with sound on) We are approaching the endgame for robotics. And this is just a beginning.
281
1,113
425
5,786
2,752,715
Great work, congrats Suning!
🤖Low-data post-training can teach a VLA policy a new robot skill. But it also makes it too attached to the training demos. We call this lock-in🔒: the policy can execute the post-training task, yet fails to respond to seemingly obvious prompt changes. DeLock preserves steerability using only the policy’s own pretrained knowledge. No extra supervision needed!🚀🚀🚀 #Robotics #AI #EmbodiedAI #VLA
1
2
813
🤖Low-data post-training can teach a VLA policy a new robot skill. But it also makes it too attached to the training demos. We call this lock-in🔒: the policy can execute the post-training task, yet fails to respond to seemingly obvious prompt changes. DeLock preserves steerability using only the policy’s own pretrained knowledge. No extra supervision needed!🚀🚀🚀 #Robotics #AI #EmbodiedAI #VLA
7
43
10
182
33,259