Han Xue (@HanXue012) profile picture

Scaling Up @liberai_robot Ph.D. student @sjtu1896 | Interesed in Robot Learning & Manipulation

Joined October 2023
Photo by Han Xue (@HanXue012)Photo by Han Xue (@HanXue012)Photo by Han Xue (@HanXue012)Photo by Han Xue (@HanXue012)Photo by Han Xue (@HanXue012)Photo by Han Xue (@HanXue012)Photo by Han Xue (@HanXue012)Photo by Han Xue (@HanXue012)Photo by Han Xue (@HanXue012)Photo by Han Xue (@HanXue012)
Reactive Diffusion Policy (RDP) is selected as Best Student Paper Finalist 🏆 in #RSS2025! I'll present RDP in tomorrow afternoon (Jun. 22) in the main conference. Welcome to chat more in the poster session!
Humans can easily perform complex contact-rich tasks with vision and touch, but these tasks remain challenging for robots. How can we resolve this from both the algorithm side and the data side? Introducing Reactive Diffusion Policy (RDP),a slow-fast imitation learning algorithm for visual-tactile/force policy learning, and TactAR, a teleoperation system with tactile/force feedback via AR. Website: reactive-diffusion-policy.gi… Arixv: arxiv.org/abs/2503.02881
3
5
71
4,821
By co-designing model,data and hardware, we have found a promising scaling path to covert web-scale human experiences into robot actions: human videos for generalization, LiberSense Glove data for action alignment, and human-like robots to reduce embodiment gap.
Introducing Liber-0, our World Action Model for zero-shot robot manipulation—including dexterous, multi-step tasks. Across 60 tasks proposed by GPT-6 and 600 trials, Liber-0 achieved 36% overall success without robot demonstrations collected specifically for those tasks: • Pick-and-place: 63.5% • Other manipulation tasks: 34.0% • Challenging multi-step tasks: 10.5% Zero-shot capability extends beyond individual actions to coordinated sequences. Pretrained on 2.2M+ hours of human experience, Liber-0 jointly predicts future observations and actions to turn human experience into robot skills. Technical blog: liberai.com/blog/liber0
2
12
818
[1/7] We trained a world model on 30,000 hours of first-person video. It learned exactly where the hands go. What happens to the things they touch, much less so. Going from 300 to 30,000 hours did not narrow the gap, and a 4× larger model barely did either.
41
61
31
475
100,492
Robot Policy Iteration is bottlenecked by the "PhD-in-the-loop" requirement for data collection. We introduce 📱RoboPocket, a portable system that turns your smartphone into an intelligent co-pilot for robot post-training. No physical robot? No problem. Improve policies instantly in the wild! 🤖 🔗Project Page: robo-pocket.github.io
4
29
4
167
17,186
RoboPocket excels in the wild. We deploy a fleet to 4 different users in 4 distinct scenes. With just 12 interactive corrections per scene, success rates jumped from ~0.50 to >0.80. This proves the scalability of distributed human-in-the-loop refinement without hardware limits. 🌍 [6/n]
2
1
8
518
Check out our full paper and videos below! Website: https://robo-pocket.github.ioArxiv: arxiv.org/abs/2603.05504Work done w/ @junjie_fang_, @wendi_chen_, @uimicro, @Tiam_Lme, Yi Wang, Yuting Zhang, @JunLv98, @ChuanWen15, and @lucewu.
1
2
6
508
Large behavior models can enable dexterous robot manipulation, but limited robot data hurts generalization. This paper studies how different types of co-training data impact policy performance and what works best.
2
13
3
146
8,337
Han Xue retweeted
Introducing DreamZero 🤖🌎 from @nvidia > A 14B “World Action Model” that achieves zero-shot generalization to unseen tasks & few-shot adaptation to new robots > The key? Jointly predicting video & actions in the same diffusion forward pass Project Page: dreamzero0.github.io/ 🧵 (1/10)
18
49
18
263
65,989
Do you ever find finetuning VLA overfits to the target task, to the point where generalist ability is lost and even minor deviations beyond the SFT data break the policy? We found an extremely simple solution: directly merge the base and finetuned policy in weight space 🤯 👇🧵
8
48
14
382
128,994
Today, we present a step-change in robotic AI @sundayrobotics. Introducing ACT-1: A frontier robot foundation model trained on zero robot data. - Ultra long-horizon tasks - Zero-shot generalization - Advanced dexterity 🧵->
427
622
455
5,350
2,060,875
Some of my favorites on reward modeling, RL, and robust VLA from the community: arxiv.org/abs/2505.10911 from @Jesse_Y_Zhang arxiv.org/abs/2510.14830 from @kunlei15 arxiv.org/abs/2509.25358 from @QianzhongChen and pi0.6 from @physical_int today: physicalintelligence.company…
17
1
112
8,640
Han Xue retweeted
RL feels messy, but a two-axis view—data source (on/off/offline) × update schedule (one-step/multi-step/iterative)—brings order. I wrote a post unifying them with shared equations. lei-kun.github.io/blogs/rl.h… Robotic FMs (e.g., GEN-0, pi_0.5) grow via a data flywheel. Best fit: multi-step updates—conservative yet exploratory—then switch to iterative RL to surpass/align human ceilings.
11
57
2
471
30,011
Real2Sim evaluation for deformable objects is very hard. Nice work from @kaiwynd!
🧵 Evaluating robot policies in the real world is slow, expensive, and hard to scale. During my internship at @SceniXai this summer, we had many discussions around the two key questions: how accurate must a simulator be for evaluation to be meaningful, and how do we get there? Our new framework, Real2Sim-Eval, takes a step toward that answer. By combining Gaussian Splatting for photorealistic rendering and soft-body digital twins for realistic dynamics, we make simulation predictive of real-world performance. 👉 real2sim-eval.github.io
1
3
868
Introducing GEN-0, our latest 10B+ foundation model for robots ⏱️ built on Harmonic Reasoning, new architecture that can think & act seamlessly 📈 strong scaling laws: more pretraining & model size = better 🌍 unprecedented corpus of 270,000+ hrs of dexterous data Read more 👇
47
277
111
1,491
501,652
Imagine a future where we have a giant robot factory. How does the robot realize it has gone off track? Do we need to equip each of them with a human operator for safety's sake? Our answer is: No! Introducing ARMADA: Autonomous Online Failure Detection and Human Shared Control Empower Scalable Real-world Deployment and Adaptation! Project Website: virlus.github.io/armada/ Paper link: arxiv.org/pdf/2510.02298 Open-source code: github.com/Virlus/armada
8
55
4
311
37,011
Can robots learn from their own experiences like humans? How can they go beyond demonstrations to discover novel behaviors? ☕️Introducing SOE — Sample-Efficient Robot Policy Self-Improvement via On-Manifold Exploration! By constraining exploration to the manifold of valid actions, SOE yields diverse yet temporally coherent behaviors — enabling structured, efficient exploration. With simple rejection sampling + imitation learning, the collected rollouts fuel powerful policy self-improvement. ericjin2002.github.io/SOE
2
27
5
132
13,873