Chaoyi Pan (@ChaoyiPan) header
Chaoyi Pan (@ChaoyiPan) profile picture

Ph.D. @CarnegieMellon @CMU_ECE @CMU_Robotics @LeCARLab | Intern @AIatMeta @Meta | @Tsinghua_Uni alum | Exploring the intersections of Learning & Control

Pittsburgh, PA
Joined July 2022
Photo by Chaoyi Pan (@ChaoyiPan)Photo by Chaoyi Pan (@ChaoyiPan)Photo by Chaoyi Pan (@ChaoyiPan)Photo by Chaoyi Pan (@ChaoyiPan)Photo by Chaoyi Pan (@ChaoyiPan)Photo by Chaoyi Pan (@ChaoyiPan)Photo by Chaoyi Pan (@ChaoyiPan)Photo by Chaoyi Pan (@ChaoyiPan)Photo by Chaoyi Pan (@ChaoyiPan)Photo by Chaoyi Pan (@ChaoyiPan)
Generative models (diffusion/flow) are taking over robotics 🤖. But do we really need to model the full action distribution to control a robot? We suspected the success of Generative Control Policies (GCPs) might be "Much Ado About Noising." We rigorously tested the myths. 🧵👇
18
89
24
550
113,300
Chaoyi Pan retweeted
Common belief: Robot policies need continuous action heads; discrete actions are 𝐥𝐞𝐬𝐬 𝐧𝐚𝐭𝐢𝐯𝐞 𝐚𝐧𝐝 𝐥𝐞𝐬𝐬 𝐩𝐫𝐞𝐜𝐢𝐬𝐞. ❗We find policies w/ the simplest discrete representations 𝐦𝐚𝐭𝐜𝐡 𝐨𝐫 𝐛𝐞𝐚𝐭 continuous ones—in sim & real, in single-step & diffusion/flow settings. Here's why 🧵
13
49
12
288
31,238
Chaoyi Pan retweeted
🤖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
33,264
Moving from scaling to research: How do we train VLAs efficiently with limited data? Haoyang’s new project shows why object-centric co-training is key. Plus, their repo is highly optimized for multi-GPU training. Great resource!
VL data co‑training is widely adopted in SoTA VLAs, but its impact is rarely isolated. Our controlled experiment reveals that auxiliary VL supervision generalizes policies to tasks seen only in VL data, even when the action head is only trained on a smaller robotic dataset.
2
6
50
6,139
🕸️ SPIDER is now OPEN SOURCE! 🕸️ The first framework for both humanoids and dexterous hands retargeting. ✅ Physics-Based: No artifacts or ghosting. ✅ General: Supports 9+ robots (G1, H1, Allegro, etc.) 📦 Repo: github.com/facebookresearch/… 📓 Tutorial: github.com/facebookresearch/…
39
1
221
16,391
Action Chunking is standard in robot learning, but we finally have a formal "why." 🔑 Key Insight: AC leverages the open-loop stability of the system. In stable dynamics (e.g., end-effector control), AC yields horizon-free imitation guarantees. Great work by @ThomasTCKZhang
🤖🤖Very excited to finally share our new work “Action Chunking and Exploratory Data Collection Yield Exponential Improvements in Behavior Cloning for Continuous Control” Everyone in robotics does action-chunking, but why does it actually work?🤔🤔And, what can theory tell us about the properties of data we should be collecting for robotic behavior cloning? 🧵1/N
1
7
1,624
🤖🤖Very excited to finally share our new work “Action Chunking and Exploratory Data Collection Yield Exponential Improvements in Behavior Cloning for Continuous Control” Everyone in robotics does action-chunking, but why does it actually work?🤔🤔And, what can theory tell us about the properties of data we should be collecting for robotic behavior cloning? 🧵1/N
5
64
7
430
73,442
If you see @max_simchowitz @GuanyaShi @nmboffi @guannanqu at NeurIPS, feel free to chat with them about our work! Work led by me w/ Giri Anantharaman, Nai-Chieh Huang, Claire Jin, Daniel Pfrommer, Chenyang Yuan, Frank Permenter, @guannanqu , @nmboffi, @GuanyaShi, @max_simchowitz
1
21
2,696
🧐🧐 Why do we pretrain LLMs with log likelihood? Why does action chunking work so well in robotics? Why is EMA so ubiquitous? And could their be a mathematical basis for Moravec’s paradox? 🤖🤖 Come check out our NeurIPS  2025 Tutorial “Foundations of Imitation Learning” with @canondetortugas and Adam Block, Tuesday 130-4pm, to find out! (🧵for details)
6
39
4
273
23,010
Another big leap for learning from humans! Amazing to see human→robot learning done with just Aria 2 glasses from @irmakkguzey. 2026 is going the year of 3D hardware + foundation models + policy learning coming together. 🤖🪞
Dexterous manipulation by directly observing humans - a dream in AI for decades - is hard due to visual and embodiment gaps. With simple yet powerful hardware - Aria 2 glasses 👓 - and our new work AINA 🪞, we are now one significant step closer to achieving this dream.
2
25
4,593
A great, down-to-earth work on tackling a locomanipulation sim2real problem. (Note to self: find 64 spare GPUs 😂) Amazing results!
Zero teleoperation. Zero real-world data. ➔ Autonomous humanoid loco-manipulation in reality. Introducing VIRAL: Visual Sim-to-Real at Scale. We achieved 54 autonomous cycles (walk, stand, place, pick, turn) using a simple recipe: 1. RL 2. Simulation 3. GPUs Website: viral-humanoid.github.io/ Arxiv: arxiv.org/abs/2511.15200 Deep dive with me: 🧵
3
15
2,400
🕸️ Introducing SPIDER — Scalable Physics-Informed Dexterous Retargeting! A dynamically feasible, cross-embodiment retargeting framework for BOTH humanoids 🤖 and dexterous hands ✋. From human motion → sim → real robots, at scale. 🔗 Website: jc-bao.github.io/spider-proj… 🧵 1/n
11
58
16
232
66,287
🧵 5/n SPIDER is also a data engine: We vary object size, states, and scenes to produce massive, physics-valid training data.
1
2
1
7
811