🤖Low-data post-training can teach a VLA policy a new robot skill. But it also makes it too attached to the training demos.
We call this lock-in🔒: the policy can execute the post-training task, yet fails to respond to seemingly obvious prompt changes.
DeLock preserves steerability using only the policy’s own pretrained knowledge. No extra supervision needed!🚀🚀🚀
#Robotics #AI #EmbodiedAI #VLA
ICRA 2026 Accepted! See you in Vienna🇦🇹! #ICRA#robotics 🤖
Can we make Diffusion Policies robust and interpretable🤔?
Introducing MoE-DP: An MoE-Enhanced architecture that decomposes complex skills to recover from failures.
👇 https://t.co/zHSdsXAOhA (1/n)
Big thanks to the incredible team for making this possible: @Shutter_Chen, TIanhai Liang, @suning_huang, Maanping Shao, Feihong Zhang, Botian Xu, @ZhengrongX, @HarryXu12
Paper link: https://t.co/70Hp9cTDrX (7/n)
How to rearrange tasks? We propose a hierarchical VLM-based planning framework:
1️⃣ Summarization: Maps experts to semantic skills (e.g., Expert 5→Close Drawer). 2️⃣ Execution: Translates high-level goals into expert sequences to steer the policy. (6/n)