CMU MSR ’27 @CMU_Robotics | SJTU ’25 @sjtu1896
Pittsburgh, PA
Joined September 2024
- Tweets27
- Following117
- Followers109
- Likes23
Pinned Tweet
FADA got into CoRL 2026, and the code is out 🎉
Adapt a humanoid to new conditions from two minutes of its own experience. No rewards, no demonstrations, no simulator retuning.
Now you can run the full pipeline yourself and adapt your own humanoid to whatever environment you drop it into, in only two minutes.
💻 github.com/LeCAR-Lab/FADA
Humanoid robots often fail when conditions change.
A heavy backpack. A soft floor. A steep slope. Suddenly, the same controller may not work.
Meet FADA.
🦾
It adapts a humanoid to new conditions using only its own experience, keeping what the robot wants to do and changing only how it does it.
About two minutes. No rewards. No demonstrations. No simulator retuning.
🧵
Angchen Xie retweeted
Impressive! Never bet against computing 🔥
A few thoughts:
1. Extending to scene- and object-interaction setting will be very interesting. Likely need to design some smart curriculum / representation to make the model robust & generalizable.
2. When MJWarp supports hybrid analytic differentiability (I guess it will come soon from my student @etaoxing. Code is already here: github.com/google-deepmind/m…), It would be interesting to compare APG via learned models v.s. APG via differentiable MJWarp.
3. Real-world fine-tuning will be super cool. I believe that is the biggest advantage of MBRL: You can fine-tune the model in the real world via supervised learning and eventually go beyond physics-based sim.
We explored it in FADA: xitter.cf/AngchenXie/status/2072…. However, FADA only fine-tunes the IDM (Inverse Dynamics Model) in a task-specific manner. How to extend it to general settings is very interesting but also challenging.
4. One of my favorite ICML 2022 papers is "Do Differentiable Simulators Give Better Policy Gradients?" from Terry Suh, @max_simchowitz, @KaiqingZhang and @RussTedrake. How exactly first-order PG outperforms zeroth-order PG in contact-rich problems is still a mystery to me.
Inspired by this paper, I think it would be cool to explore some hybrid zeroth-order and first-order methods, and even do some "annealing" (smooth the dynamics in early-stage training and gradually switch to full dynamics). We found it critical in zeroth-order optimal control in DIAL-MPC xitter.cf/HaoruXue/status/183901… and I am very curious if APG will have similar conclusions.
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/
Angchen Xie retweeted
RPL code is released at github.com/amazon-far/RPL, including experts training, multi-depth distillation and sim2sim. If you have an Xbox controller lying around, it’s a lot of fun playing it in Mujoco.
Its perceptive features were adopted by some cocurrent work like LadderMan (ladderman-robot.github.io/) and PHP (php-parkour.github.io/) with promising results.
RPL was also accepted to CoRL2026. See you in Austin.
Robust humanoid perceptive locomotion is still underexplored. Especially when different cameras see different terrains, paths get narrow, and payloads disturb balance...
Introduce RPL, tackling this with one unified policy:
• Challenging terrains (slopes, stairs and stepping stones);
• Multiple directions;
• Payloads;
Trained in sim. Validated long-horizon in the real world.
Watch the robot walk it all🦿
Details below👇
Angchen Xie 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 🧵
Angchen Xie retweeted
Super excited to share some of the work I’ve been doing at the GEAR lab @NVIDIAAI.
After three months, we’re reintroducing ASPIRE: Agentic Skills Discovery for Robotics (v1.0). Enough of Astra ✨ doing pick-and-place—let it learn reusable strategies and solve a Rubik’s Cube online.
Quick 🧵 for details.
Angchen Xie retweeted
Never thought of this would actually happen to me. FSD avoided the deer at ~70mph. Pretty impressive how quickly it reacted with the evasive maneuver. Thanks @Tesla
Reminds me of my research in manipulation. Policy means nothing when your robot can't execute fast enough.
Angchen Xie retweeted
🤔 Why do learning curves follow power laws—and when can training schedules change them?
🚀 We connect spectral structure to joint LR–batch schedules, deriving 3+3(+2) map and predicting (300M) LLM loss across schedules.
With @Fanghui_SgrA & Yudong Chen
arxiv.org/abs/2609.40148
Angchen Xie retweeted
Inspired by recent thoughts from senior researchers whom I deeply respect (e.g. @JitendraMalikCV @Michael_J_Black @Ken_Goldberg @phillip_isola), I wrote down some thoughts as a junior researcher too.
The recent Astra demos in dexterous manipulation are truly impressive. But they also raise two questions that I don’t think academia has good answers to yet:
- How should credit be assigned when an agent synthesizes many prior works into a new research result?
- How can academia attract and retain talents when so much of its incentive structure relies on credit?
I reflect on these questions in more detail in a blog post. Curious to hear what others think, and debates are very welcome!
toruowo.github.io/blog/posts…
Angchen Xie retweeted
🥳Introducing SNAP3D — physically compatible 3D part generation from a single image.
Existing methods can generate visually appealing 3D parts, but the generated parts are often not physically compatible for assembly.
With SNAP3D, the parts are designed to fit together in the real world.
No glue. No screws. Just print, snap, and build!
🖨️ Try it with your 3D printer!
🔗 Project page: lucytuan.github.io/SNAP3D/
💻 Code: github.com/LucyTuan/SNAP3D
🎬 Full video: youtu.be/22yRDaWabwA
Angchen Xie retweeted
Manipulation is harder when the robot flies. 🚁
Moving the arm can shift the drone, and wind can push it off course. The policy and controller have to work together.
Here's AM-Bench: A Simulation Suite and Benchmark for Aerial Manipulation
ambench.github.io
#CoRL2026
Angchen Xie retweeted
It’s great to see more people openly discussing the true limitations of their own work.
No one understands the strengths and weaknesses of your method better than you do. That puts you in the best position to contribute to principled research—and to help others build on it.
🧵
🎉 FADA is now accepted to CoRL 2026!
Along with the code, we’re releasing something a little different: a timeline of the research journey behind FADA — the ideas we tried, what failed, the choices we made, and the lessons we learned along the way.
lecar-lab.github.io/FADA-hum…
Research papers are great at telling us what worked. They are much worse at telling us everything researchers tried before getting there.
We think exposing that trajectory can give the community a much richer and more honest ground-truth context particularly in the age of AI-assisted robotics research.
We’d love to see more projects share their research journeys this way: turn your update slides to a shareable timeline!!.
Understanding why certain choices were made (and which mistakes didn’t work) is just as valuable as the final result.
Inspired from @breadli428's efforts on adding limitation threads to his works and @TairanHe99's timeline videos in VIRAL.
🧵
Nine months of work went into FADA. One paper came out.
Most of what we learned about adapting humanoids to new environments never made it into the paper: the approaches that failed, why they failed, and what they taught us about where adaptation should happen in the first place.
One failure that stuck with me: a fine-tuned dynamics model that predicted the world better but controlled the robot worse. It took us a while to accept that better prediction is not better control.
We wrote it all down. If you work on domain adaptation or sim2real, this is the part I'd want you to read 👇
🎉 FADA is now accepted to CoRL 2026!
Along with the code, we’re releasing something a little different: a timeline of the research journey behind FADA — the ideas we tried, what failed, the choices we made, and the lessons we learned along the way.
lecar-lab.github.io/FADA-hum…
Research papers are great at telling us what worked. They are much worse at telling us everything researchers tried before getting there.
We think exposing that trajectory can give the community a much richer and more honest ground-truth context particularly in the age of AI-assisted robotics research.
We’d love to see more projects share their research journeys this way: turn your update slides to a shareable timeline!!.
Understanding why certain choices were made (and which mistakes didn’t work) is just as valuable as the final result.
Inspired from @breadli428's efforts on adding limitation threads to his works and @TairanHe99's timeline videos in VIRAL.
🧵
Angchen Xie retweeted
Reality of mobile manipulation: 1,000 demos buy you 1 task, in 1 scene.
Introducing 𝐖𝐀𝐍𝐃𝐀: a synthetic data engine that turns ONE demo into thousands of trajectories across unseen scenes — and the policy works in the open world.
One demo in. An open world out.
🧵
Angchen Xie retweeted
🤖 Robots are usually taught to avoid contact unless it’s at the end effector. But caring for a person, such as rolling them over or repositioning a limb, requires embracing contact along the entire arm. How do we enable robots to perform such contact-rich tasks?
Introducing TACTIC: Tactile and Vision Conditioned Contact-Centric Control for Whole-Arm Manipulation.
#RSS2026 🧵👇
Angchen Xie retweeted
🤖Humanoid robots, to be truly useful, must handle dynamics they never saw during training: payloads, slopes, soft terrain, and more..
But these are exactly the settings where the familiar sim-to-real pain also shows up, and every humanoid researcher has probably at least once said:
“but it worked perfectly in simulation!!”
For precise whole-body control, the challenge is not just avoiding failure. Small execution errors can quickly cascade into large deviations.
In FADA, we study how humanoids can adapt to these unseen dynamics. Our key observation is simple: the robot’s intent often transfers, but the execution does not.
1/N🧵
Humanoid robots often fail when conditions change.
A heavy backpack. A soft floor. A steep slope. Suddenly, the same controller may not work.
Meet FADA.
🦾
It adapts a humanoid to new conditions using only its own experience, keeping what the robot wants to do and changing only how it does it.
About two minutes. No rewards. No demonstrations. No simulator retuning.
🧵
That's FADA.
📄 Paper: arxiv.org/abs/2606.28476
🌐 Website: lecar-lab.github.io/FADA-hum…
💻 Code: Coming soon
Built at @LeCARLab @CMU_Robotics ❤️
Couldn't have done this without this incredible team. Huge thanks to @NikhilSoban353, Ishayu Shikhare, Alan Wang, my advisors @max_simchowitz and @GuanyaShi, and special thanks to @yxyang1995, whose insight shaped a key part of this work. 🙏