Yutang Lin

Senior undergrad @PKU1898.
Robot Learning, humanoids.
Applying for PhD '27 Fall. .

[Yutang Lin @yutang_llin]

How can we represent whole-body interaction for humanoid robots—compactly enough for low-frequency planning, yet richly enough to preserve high-frequency feedback control?

Introducing FIBER! 🤖

FIBER learns a unified latent interface between high-level planning and whole-body control, enabling a humanoid to acquire diverse interaction behaviors from heterogeneous experts through a pretrained, frozen Behavior Foundation Model (BFM).

With a single planner and BFM, we demonstrate heavy-door pushing, box carrying, object grasping, controllable navigation, and more!

🌍 Website & Live Demo: https://t.co/h14roShdvb

📄 Paper: https://t.co/QBjYNLCzEW

A thread on the ideas behind FIBER 🧵👇

[Yutang Lin @yutang_llin]

@Modern_Gangster In our test sonic's hidden space is not good enough for projection, even without FSQ discretization.

[Yutang Lin @yutang_llin]

How can we represent whole-body interaction for humanoid robots—compactly enough for low-frequency planning, yet richly enough to preserve high-frequency feedback control?

Introducing FIBER! 🤖

FIBER learns a unified latent interface between high-level planning and whole-body control, enabling a humanoid to acquire diverse interaction behaviors from heterogeneous experts through a pretrained, frozen Behavior Foundation Model (BFM).

With a single planner and BFM, we demonstrate heavy-door pushing, box carrying, object grasping, controllable navigation, and more!

🌍 Website & Live Demo: https://t.co/h14roShdvb

📄 Paper: https://t.co/QBjYNLCzEW

A thread on the ideas behind FIBER 🧵👇

[Qiayuan Liao @qiayuanliao]

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: https://t.co/6gc9WKm9qY
Demo: https://t.co/OHRvx5ogQY

[Zihan Wang @Z1hanW]

Humans generalize from rich interaction experience. We bring this to humanoid loco-manipulation through scalable real-to-sim-to-real.

Introducing PRISM, we expand 4 real videos into 256 counterfactual variants, recover robot-object trajectories, and train one policy that generalizes across diverse objects, scales, and spatial layouts.

CoRL’26 🚀 Website for details 👉 https://t.co/DUf2DIrvai

Work done at Amazon FAR w/ @zhenkirito123 @pabbeel @rocky_duan @JitendraMalikCV @carlo_sferrazza @ckarenliu @GuanyaShi @akanazawa

[Yen-Jen Wang ✈️ CoRL @wangyenjen]

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

[Yutang Lin @yutang_llin]

5/5 On a personal note, this is also my first paper as the sole student author!

I'm incredibly grateful to my mentors for their patient guidance and support over the past several months. I've learned so much throughout this project, and it has genuinely changed how I think about and approach research.

The code is still being organized, but detailed training configurations and architectural details are available in the appendix.

Thanks for reading, and I'd be happy to discuss any questions or ideas!

P.S. There's also an interactive live demo on our website, although the INT8-quantized planner can be a little unpredictable :)

[Yutang Lin @yutang_llin]

4/5 A few additional technical insights from this project:

(1) Multi-γ Forward-Backward Learning

Our hypothesis is that FB representations primarily capture long-horizon future occupancy, while precise tracking, especially for highly dynamic body parts, requires stronger short-horizon reachability information.

We explore a dual-γ FB objective to balance motion representation capacity and tracking precision.

(2) Interaction-Preserving Skill Composition

Given a BFM and an interaction task, can we efficiently reuse the BFM's native capabilities without destroying the interaction?

We explore successor-feature-guided RL to learn compositions such as carrying a box while walking, crouching, or kneeling, while maintaining the original interaction.

[Yutang Lin @yutang_llin]

3/5 What can we achieve with this representation?

We train a single planner conditioned on perception and language commands, which predicts latent commands executed by the same frozen BFM.

Despite the diversity of the underlying experts, FIBER can perform a wide range of whole-body interactions on a real humanoid, including:

🚪 Pushing open heavy doors with whole-body coordination

📦 Carrying and transporting boxes while maintaining stable interactions

🦾 Object grasping and manipulation with Dex3 hands

We also automatically derive waypoint labels from the training data, allowing the planner to perform controllable transportation and steering, including navigating under obstacles by crouching.

Beyond individual tasks, we evaluate skill chaining in MuJoCo, where the humanoid sequentially opens a door, carries a box, and returns to its starting position.

All of these behaviors are realized through the same latent planning and control interface.

[Yutang Lin @yutang_llin]

2/5 Our core idea is to consolidate interaction experts in the latent space of a pretrained, frozen BFM.

Given an expert's state-action trajectory (s, a), FIBER performs action-aware inversion: we optimize a sequence of latent commands z such that the BFM, conditioned on z at each expert state, produces feedback actions that closely match those of the expert.

The distinction here is important: we are not simply encoding reference motions into a latent space.

For interaction tasks, reproducing the same motion is not necessarily sufficient. What matters is how the robot generates actions in response to its current physical state—how the interaction is actually realized.

By aligning the BFM's local feedback with that of the expert, we aim to preserve this interaction realization within a shared latent representation.

We can then train a unified planner directly over these latent sequences, instead of learning separate control policies for every interaction task.

[Yutang Lin @yutang_llin]

@Haian_Jin Great work, congrats!

[Yixuan Wang @YXWangBot]

1/ World models are getting popular in robotics 🤖✨
But there’s a big problem: most are slow and break physical consistency over long horizons.

2/ Today we’re releasing Interactive World Simulator:
An action-conditioned world model that supports stable long-horizon interaction.

3/ Key result:
✅ 10+ minutes of interactive prediction
✅ 15 FPS
✅ on a single RTX 4090🔥

4/ Why this matters: it unlocks two critical robotics applications:
🚀 Scalable data generation for policy training
🧪 Faithful policy evaluation

5/ You can play with our world model NOW at https://t.co/SBqVDzYn86. NO git clone, NO pip install, NO python. Just click and play!

NOTE ⚠️
ALL videos here are generated purely by our model in pixel space! They are **NOT** from a real camera

More details coming 👇 (1/9)

#Robotics #AI #MachineLearning #WorldModels #RobotLearning #ImitationLearning

[Yutang Lin @yutang_llin]

@pablovelagomez1 @rerundotio @NVIDIAAIDev Looks amazing, we definitely need something like this.

[Yutang Lin @yutang_llin]

@Pinyuan3 @adeli_hossein @wenxuan_guo_ @LibraCheng @ethanhwang_ @KriegeskorteLab Interesting, congratulations!

[Yutang Lin @yutang_llin]

@robotsailor lol, so true

[Jinrui Han @jinrui82]

Introducing our new project——HUSKY: Humanoid Skateboarding System via Physics-Aware Whole-Body Control ！🎉🎉
Powered by mjlab framework, humanoid skateboarding is coming true right now ! 🙌
View our work at https://t.co/tlHETwj1Jr.
Paper available at https://t.co/SO29rGw3g0.

[Jesse Zhang @Jesse_Y_Zhang]

A reward model that works, zero-shot, across robots, tasks, and scenes?

Introducing Robometer: Scaling general-purpose robotic reward models with 1M+ trajectories.

Enables zero-shot: online/offline/model-based RL, data retrieval + IL, automatic failure detection, and more!

🧵 (1/12)

[Yutang Lin @yutang_llin]

Six months ago, I watched Flow Matching doing push-ups and flips—now it’s evolved. It’s one thing for generative models to provide the raw power, but seeing people squeeze that massive, multi-step denoising policy into a Jetson Orin on the G1 is next-level engineering lol. Absolutely wild.

[Siyuan Huang @siyuanhuang95]

You might have seen the WuBOT performing at the 2026 Spring Festival Gala; however, most high-dynamic extreme motions you see are executed by overfitted tracking policies. Until now, training a unified policy capable of performing various extreme motions with a high success rate remained an unsolved challenge.

We spent an entire year digging into the barrier between general tracking and extreme physical behaviors. After burning through dozens of G1 robots, we finally identified the bottleneck of learning and physical executability.

With these discoveries, we developed OmniXtreme: the first general policy that can execute diverse extreme motions, including consecutive flips, extreme balancing, and even breakdancing with rapid contact switches!

This capability is achieved by pre-training a flow-based generative control policy and then post-training with actuation-aware residual RL for complex physical dynamics—a step we found critical for successful real-world transfer.

This work is a joint collaboration with @UnitreeRobotics. Together, we are pushing the physical limits of humanoid robots. It is incredibly exciting to see a general "robot gymnast" and "robot breakdancer" come to life! It was also our first time publishing a paper with XingXing, which was an enlightening experience.

The model checkpoints are now released—we welcome you to play with them! 📦

📄 Paper: https://t.co/ySKLac8w6i
🌐 Project: https://t.co/ortlLW24JB
💻 Code: https://t.co/IJ4XHYT4Qj

[Yutang Lin @yutang_llin]

Impressive!!! Reducing highly constrained reference motion input to simple velocity commands is a big step for such contact-rich and interactive task.

[Zhen Wu @zhenkirito123]

Can humanoids perform agile, autonomous, long-horizon parkour—based on what they see in the world?
We present 𝗣𝗲𝗿𝗰𝗲𝗽𝘁𝗶𝘃𝗲 𝗛𝘂𝗺𝗮𝗻𝗼𝗶𝗱 𝗣𝗮𝗿𝗸𝗼𝘂𝗿 (𝗣𝗛𝗣): a framework that chains dynamic human skills using onboard depth perception for long-horizon traversal.

1/6

[Yuanhang Zhang @Yuanhang__Zhang]

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👇

[Zhen Wu @zhenkirito123]

Can humanoids perform agile, autonomous, long-horizon parkour—based on what they see in the world?
We present 𝗣𝗲𝗿𝗰𝗲𝗽𝘁𝗶𝘃𝗲 𝗛𝘂𝗺𝗮𝗻𝗼𝗶𝗱 𝗣𝗮𝗿𝗸𝗼𝘂𝗿 (𝗣𝗛𝗣): a framework that chains dynamic human skills using onboard depth perception for long-horizon traversal.

1/6