Zhecheng Yuan (@fancy_yzc) header
Zhecheng Yuan (@fancy_yzc) profile picture

@PokeBot_Global. Final year PhD Tsinghua University, IIIS.

Joined July 2021
Photo by Zhecheng Yuan (@fancy_yzc)Photo by Zhecheng Yuan (@fancy_yzc)Photo by Zhecheng Yuan (@fancy_yzc)Photo by Zhecheng Yuan (@fancy_yzc)Photo by Zhecheng Yuan (@fancy_yzc)Photo by Zhecheng Yuan (@fancy_yzc)Photo by Zhecheng Yuan (@fancy_yzc)Photo by Zhecheng Yuan (@fancy_yzc)Photo by Zhecheng Yuan (@fancy_yzc)Photo by Zhecheng Yuan (@fancy_yzc)
๐ŸŒถ๏ธ๐Ÿค–๐–๐ž ๐œ๐จ๐จ๐ค๐ž๐ ๐ญ๐ก๐ž ๐ซ๐จ๐›๐จ๐ญ๐ข๐œ๐ฌ ๐ฐ๐จ๐ซ๐ฅ๐โ€™๐ฌ ๐Ÿ๐ข๐ซ๐ฌ๐ญ ๐ฉ๐ฅ๐š๐ญ๐ž ๐จ๐Ÿ ๐Œ๐š๐ฉ๐จ ๐“๐จ๐Ÿ๐ฎ.๐Ÿค–๐ŸŒถ๏ธ This was not just a cooking demo. It was a 30-minute, long-horizon robotics challenge packed with delicate, continuous, high-dexterity actions. With only a small number of demonstrations, our model learned to perform complex manipulation over an extended sequence. At the same time, we pushed hard on motion control and system-level optimization, making the robot move smoothly while keeping the high success rate. For us, this plate of Mapo Tofu is more than a dish. It is a small but meaningful step toward home robots that can handle real everyday tasks with the dexterity, consistency, and reliability people expect.
17
26
9
147
26,670
Super reactive!๐Ÿ˜ฎ
Introducing โšก๏ธFlash ALOHA: Learning Fast and Precise Manipulation! With just 50 demonstrations, Flash ALOHA learns to autonomously complete dynamic and precise tasks at near human speed. Everything open sourced! (1/11) Team: @Qi_Wu577 @YihuaiGao @chelseabfinn @zipengfu
1
5
784
Amazing work Zhecheng!
Reliable real-world manipulation requires policies that can improve from deployment experience. Introducing PIVOT: Policy Improvement via Trust-Region Optimization from Deployment Experience โ€” 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
3
297
Zhecheng Yuan retweeted
Deployment time is a surprisingly revealing metric for exposing the limitations of a policy model. Honestly, I feel like "Recursive Self-Improvement" is sometimes old wine in a new bottle โ€” essentially DAgger, real-world RL, and iterative policy refinement under a fancier name.
Reliable real-world manipulation requires policies that can improve from deployment experience. Introducing PIVOT: Policy Improvement via Trust-Region Optimization from Deployment Experience โ€” 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
1
6
437
Zhecheng Yuan retweeted
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 โ€” 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
Zhecheng Yuan retweeted
Impressive real-world results
Reliable real-world manipulation requires policies that can improve from deployment experience. Introducing PIVOT: Policy Improvement via Trust-Region Optimization from Deployment Experience โ€” 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
2
38
4,248
There r lots of positive signals from negative outcomes (ie. learning from mistakes). Congrats @fancy_yzc & team on the cool work!
Replying to @fancy_yzc
4/8 Learning from negative experience makes a measurable difference. Excluding negative-advantage samples reduces zip-tie success from 86% to 60%. Removing the trust-region constraint causes training to diverge. With the same collected experience, TRFO also outperforms HG-DAgger and RECAP on laundry folding and zip-tie threading. Both the negative learning signal and the constraint on its strength contribute to effective policy improvement.
1
3
11
672
Reliable real-world manipulation requires policies that can improve from deployment experience. Introducing PIVOT: Policy Improvement via Trust-Region Optimization from Deployment Experience โ€” 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.
6
27
5
188
41,327
8/8 Multi-task scalability: PIVOT also supports joint post-training across multiple skills. A single policy trained on laundry, zip-tie, and sachet tasks achieves: โ€ข67.2 T-shirts/hour โ€ข87% zip-tie success โ€ข78% sachet success Performance is comparable to separate task specialists, combining rapid manipulation and millimeter-level precision within one model.
1
4
527
We will present PIVOT at CoRL 2026. ๐Ÿ“ Poster Session 1 ยท Poster #84โ€จ๐Ÿ—“ Monday, November 9, 2026 ยท 5:00โ€“6:30 PM CSTโ€จLone Star, 3rd floor, JW Marriott Austin Please stop by to discuss robot post-training, learning from deployment experience, and reliable manipulation. Many thanks to my collaborators Shenghan Lin @ShenghanLin2006 , Qicheng Xu @HACLINExk, Xiangchen Tian, Chenhao Lu @chenhao_lu49219 , Zilin Kang @zilin_kang, Muhan Wang, Xiuyu Xie, and Huazhe Xu @HarryXu12 for their contributions to this work. Webpage: linshey.com/pivot/ Paper link: openreview.net/pdf?id=2vx7I6โ€ฆ
1
7
584
Zhecheng Yuan retweeted
Just started open-sourcing ART-Glove! github.com/linchangyi1/ART-Gโ€ฆ Mechanical files and assembly guide are available now. PCB, FPCB, code, and video guide are coming next! Iโ€™ll also bring the glove to the CoRL demo session on Monday. Come check it out!
Manipulation happens through surfaces. To understand contact-rich dexterous interaction, motion alone is not enough. We also need to know surface properties and contact state. Excited to share ART-Glove, an articulated tactile glove that captures contact-grounded information while preserving human dexterity. It provides: - Known Geometry: 16 rigid functional surfaces - Surface Motion: 22 anatomically aligned joints - Tactile Contact: 2048 piezoresistive taxels Huge thanks to my advisor Ding @zhao__ding, and to Yuxiang @yxyang1995, Maria @bauzavillalonga, Marissa, and Peide @peide_huang for the valuable advice and discussions. Paper: arxiv.org/abs/2606.16370
10
21
3
177
8,363
Kinematic retargeting can produce similar motion, but once object interaction is involved, tiny inaccuracies get amplifiedโ€”it canโ€™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 demonstrationsโ€”with human data only? Introducing TouchBridge: transferring contact-rich skills from human hands to robot hands through touch. ๐Ÿ–๏ธ๐Ÿค– ๐Ÿ“„ arXiv: arxiv.org/abs/2610.40349 ๐ŸŒ 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 7.0ร— 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 bulbโ€”hereโ€™s how it works. ๐Ÿงต๐Ÿ‘‡
1
1
3
315
Kinda scaryโ€ฆ
Replying to @Lingxiao234
6/n There are more exciting opportunities beyond this: 1. Agent for robot learning science. Agents may be better than us at understanding these models, and unlike us they can carry what they learn about one architecture into another โ€” bridging the science of AI across domains inside one model. 2. A new way for agentic robotics: agentic surgery on VLA/WAM, instead of retraining them. 3. Agentic few-shot policy adaptation, like sim2real adaptation.
1
1
415
interesting!
We test Astra weight edit on three robot domains: VLA, WAM and humanoid control. On one task it edited only ๐Ÿ’ ๐ฉ๐š๐ซ๐š๐ฆ๐ฌ ๐จ๐ฎ๐ญ ๐จ๐Ÿ ๐Ÿ‘.๐Ÿ”๐ for ฯ€0.5, and success went ๐Ÿ๐Ÿ–/๐Ÿ—๐Ÿ โ†’ ๐Ÿ–๐Ÿ/๐Ÿ—๐Ÿ on held-out unseen episodes. How Astra got there actually interests me more ๐Ÿ‘‡
1
347
Zhecheng Yuan retweeted
Marin 535B-A23B now on 65K contexts... Counting down...
11
25
6
405
32,386