Jinliang Zheng

PhD student at @air_tsinghua

[Jinliang Zheng @2_toinf]

@LeRobotHF Many thanks to the LeRobot team for their ongoing support. Give X-VLA a try and let it spark your own ideas!

[LeRobot @LeRobotHF]

🚀 Introducing X-VLA ; LeRobot’s new soft-prompted Vision-Language-Action model.

X-VLA is built to scale across many embodiments: different robots, cameras, action spaces, and environments, all handled by one unified transformer backbone.

- Generalist across robots (Franka, WidowX, Agibot, sim + real)
- Soft-prompt domain IDs let the model adapt to new hardware with tiny learnable embeddings
- Flow-matching + transformer core for smooth, continuous 50 Hz control
- Pretrained on a mixed-embodiment dataset spanning 7+ platforms and diverse tasks
- Fine-tune on any dataset using one of the 6 checkpoints we provide out of the box.

[Jinliang Zheng @2_toinf]

🚀 New: X-VLA (0.9B)-A Lightweight VLA model for cross-embodiment manipulation
🧺 2+ hrs cloth folding autonomy, trained with only 1.5K demos
📊 Benchmarks: 96% Simpler-WidowX, 97.4% Libero-Long, 70% RoboTwin, + more
Project Page: https://t.co/Jq3e66ypQs
(1/N)

[Jinliang Zheng @2_toinf]

Neat and scalable architecture design, together with well-ablated training recipes, will be released — fully transparent, with no hidden tricks.
(2/N)
📄 Paper: https://t.co/DO5LGc4IBm
📷 Code: https://t.co/4e8HpyQPCP

[Jinliang Zheng @2_toinf]

🎉 Announcing Soft-FOLF — the first open-source, high-quality cloth-folding dataset.
✨ Features:
• DAgger-style data collection
• 1K+ episodes (~24h)
• Diverse cloth targets
🚀 Powered X-VLA’s 2+ hr autonomous folding.
📂 Dataset coming to HF in ~1 month.
(5/N)
Demo:

[Jinliang Zheng @2_toinf]

Among the most comprehensive evaluations of VLA models on simulation benchmarks so far — showing amazing, state-of-the-art performance across the board🥳.
(4/N)

[Jinliang Zheng @2_toinf]

🚀 Exciting scaling dynamics emerge when X-VLA is pretrained on heterogeneous dataset mixtures.
(3/N)

[Zheng Yinan @ZhengYinan2001]

🥳New SOTA on nuPlan! #ICLR2025
🚗Diffusion Planner leverages diffusion models with a specifically designed architecture for high-performance motion planning, reducing dependence on rule-based refinement.
🔗 https://t.co/pILK7RN65w

[Jinliang Zheng @2_toinf]

(On going) More embodiments, such as bimanual robots, mobile robots and autonomous driving cars, will be absorbed into our Universal Action Space.

[Jinliang Zheng @2_toinf]

#Robotic_Foundation_Models #Universal_Actions
0.5B beats 7B models🦾!
We introduce a highly effective pretraining paradigm for embodied foundation model, powered by a Universal Action Space that boosts cross-embodiment data sharing and generalization!
https://t.co/mV5al5woGN

[Jinliang Zheng @2_toinf]

(4/n) We observe fascinating consistency in behaviors for the same code across different robots, showcasing strong cross-embodiment generalization.