I've been working on deformable object manipulation since my PhD. It was totally a nightmare years ago and my PhD advisor was telling me not to work on it for my own good. Today, at ByteDance Seed, we are dropping GR-RL, a new VLA+RL system that manages long-horizon precise... dexterous manipulation of deformable objects. This is probably the first real-world RL system to make a robot: ✅ Lace up your shoes end to end ✅ Hit millimeter tolerance repeatedly ✅ Recover from mistakes (See video!) ✅ And complete continuous shoelace threading on a real bimanual platform 📈 Success rate: ↑ from 45.7% → 83.3% Yes, robots can now actually do this. Project page: ArXiv:show more

Why “shoelace threading” matters 🤔 This task is probably one of the most challenging household robotics tasks in terms of precision: 💥 Soft-body chaos – laces deform every frame 💥 Millimeter precision – 1–2 mm slip = total failure 💥 Long-horizon manipulation – hundreds of steps where slightest drift compounds. 💥 Requires recovery – humans retry; IL models freeze It's the perfect torture test for VLA precision. Even our previous model, GR-3 — already trained on massive robot trajectories + human teleop demos + public image-text corpora — failed to get reliable policies here. So we found the bottleneck.

Two killers of imitation learning (IL): (1) Human demos are NOT optimal Humans hesitate, retry, fix mistakes mid-trajectory. IL blindly copies ALL of it — including the bad parts. (2) Training vs Deployment Misalignment VLA models output actions. To prevent jitter, robots execute post-processed versions (RHC, smoothing, ensembling). Your “predicted action” ≠ your “executed action.” On millimeter tasks, this mismatch is fatal. So IL alone was a dead end.

The Idea: If imitation is broken, then: Let the robot learn from its own experience. GR-RL = ⭐️ Offline RL (data filtering) ⭐️ Symmetry augmentation ⭐️ Online closed-loop Real-World Reinforcement Learning All on top of a single VLA foundation model. Just RGB, proprioception, and language instructions.

Offline Stage — Filter the human flaws We train a Critic Transformer via distributional RL: ⭐️ Detects “value drops” when the operator hesitates or messes up ⭐️ Slices every trajectory into high-value vs low-value segments ⭐️ Retains only the cleanest expert behavior Effect: GR-3 45.7% → 61.6% success just from removing “bad micro-behaviors.”

Morphological Symmetry Augmentation Our bi-manual robot is left–right symmetric. So we mirror EVERYTHING: 🚀 RGB 🚀 Proprioception 🚀 Actions 🚀 Language Instructions Data size doubles. Spatial reasoning robustness skyrockets. → 72.7% success.

Online Stage — Real-World Steering RL Now the robot learns ON THE PHYSICAL PLATFORM. But direct exploration in joint space causes dangerous jitter and is inefficient — you need millimeter accuracy. So GR-RL explores in the latent noise space: A tiny 51.5M-param noise-predictor nudges the flow model’s latent action distribution toward high-value regions. We also introduce a dual-buffer mechanism: 🚀 Off-policy buffer (old rollouts) for the critic warm-up 🚀 On-policy buffer (last 2 checkpoints) for stable improvement 🚀 1:1 sampling stabilizes RL and prevents catastrophic drift. Just 150 online rollouts → +10% performance boost.

The Result: Final performance: 83.3% success over continuous shoelace threading. The surprising part? GR-RL learns to: 🔥 retry when the lace slips 🔥 reposition the lace when the initial pose is bad 🔥 "self-correct" mid-task instead of freezing This is the behavior you EXPECT from embodied intelligence — not just mimicry, but competence.

GR-RL proves something: IL is inherently limited, and we can do things previously thought impossible by purely visuo-motor control simply making it RL. The future direction is clear: distill RL-enhanced behavior back into the foundation VLA, forming a self-improving, ever-growing loop. Kudos to the wonderful team! @yunfeili98, @simanbulk, @yu_cui123, Zhongren Cui, Zhigang Han, @litchiyoyo, @KT892790499, @YuxiaoLiu4, Hao Niu, Wanli Peng, Jingchao Qiao, Zeyu Ren, Haixin Shi, @ZhiSu22, Jiawen Tian, Yuyang Xiao, Shenyu Zhang, Liwei Zheng

Wow! I love how it naturally pulls the shoe closer so that it can complete the task easier. Awesome work!

haha, thanks Perry! btw congrats on your recent release on ACT-1 and Memo! Great work! I really like the design and it's amazing how much it has achieved with pure UMI data.

This is brilliant Ma

incredible work, can't wait to see more in the future.

Congrats!

@Scobleizer This is amazing

Great team! Great Work! Can not wait to see your future project! Congrats @yunfeili98 @yusufma555

This is incredible work. Kudos for seeing this work through for so many years, seems like it’s paying off!

Can you implement an aspect of leaning against the object that you're working on? This is what humans do when we tie our shoe, we don't float in space. This means that you don't have to have accurate control, you know that if you rest your finger against the shoe, then the string will be in the right place.

Hey Robert, this is a perfect point. Our model actually sometimes does it. During the offline data collection stage, most of the human demonstrators love to "float" because it is easier for them to control the robot when they became very good at it. However, this causes the issue you mentioned. But there were indeed some samples that the demonstrators love to lean on the shoes. As a result, our VLA model will demonstrate multi behavior modes, including the one you mentioned. But it is just less often, due to the initial data distribution was dominated by the other behavior mode.

Awesome, will this methodology ultimately require the robot to be trained on every task that it carries out or do you have some requirement for a "general" intelligence that can carry our arbitrary tasks? I know this is easier said than done, but interested in your opinion as I have no experience with robotics.

This task seems very difficult. Can the policy deal with different shoes?

thanks! Currently we tested on three different colored shoes. However, some other shoe designs are extremely difficult to thread, even for human demonstrators. There is much more we can do in the future :)

very impressive :)

😁 this why a professor is just person who memorized a book to tell you to read the same book. Only real world grit and experimentation is the path to success. Great Job now level up send your gear to the reserve to see what really breaks.

Really really cool

Very cool

Wow, Yusuf, that’s incredible progress from ByteDance!

Cool, congrats on the release!!

Thank you Younggyo!

Incredibly impressive, well done!

Can you clarify 1) the number of examples to reach your quoted accuracy 2) how many distinct tasks can the model remember and how does it select them

Here is prior art shoelace tyeing What matter is to advance few shot, max task number and length.

@grok analyse et explique ce qui se passe dans cette vidéo

Cool ！

还有哪些常见的日常活动对机器人而言与系鞋带平级或者更复杂？

So sick! Curious what cameras are mounted onto the end effectors?

thanks! We have discussed our hardware setups in the section 4 of our paper. It is a realsense D405 camera.

It's just a matter of time we'll solve dexterity. This is just unreal.

Amazing project! 👏 Message me directly

Awesome 👏 Message me directly

this is awesome! 🔥 congrats! 👏🏼 i noticed even though you had RGB-D cameras, you only kept the RGB for training and inference. did you find that depth data won't help here much?

This is more impressive than dancing or kung fu. Do you think it could tie a neck tie without choking a person? That's one of my benchmarks for when robots are ready for home use.

Incredible progress — deformable object manipulation has always been one of the toughest challenges in robotics, and seeing real-world RL push success rates this high is remarkable. Huge step forward for dexterous autonomy. 🤖🧠 #Robotics #ReinforcementLearning

This is pretty cool man. I'd love to learn more about your research and ByteDance Seed. Can you open your DM's? let's chat!

Impressive Stuff!!

you are the best 😍🧡💙💚 !