Suning Huang

🤖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

Great work! Congrats, Suning!

Ever fine-tuned a VLA policy on a small demo dataset and it suddenly stops listening to new instructions? This paper calls it lock-in. The model just repeats what it saw during training like always picking bread even when you say apple Low-data post-training quietly kills steerability The fix? DeLock is surprisingly simple and clever

Thanks for the thoughtful point! DeLock is not meant to replace SFT or make arbitrary unseen skills work out of the box. It aims to reduce the combinatorial burden of SFT by leveraging the pretrained backbone to connect post-trained skills with related novel instructions, so we don’t need demos for every variation. So its effectiveness depends on both the similarity between the trained and novel tasks, and how much the VLA backbone already knows about the relevant concepts/skills.

@suning_huang de-locking sounds nice but i bet it still crashes if you ask the robot to hold a coffee instead of a wrench

🤖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

Great work, congrats Suning!

We still know so little on how to use the learned representations present in VLMs for VLA training. Great work @suning_huang and team!

How well to VLAs generalize to new prompts after SFT? If you've worked with them, you'll know the answer. The problem is the fine tuning methodology, not the model. Suning has a clever and effective solution that requires no new data, just better SFT and inference methods. 👇

I am surprised that so many pre-trained knowledge can be preserved with no additional data if you finetune VLA in a proper way! Check this solid work from Suning!

Really nice work on tackling “lock-in” in VLA policies! VLA post-training robustness is a bottleneck, and it’s great to see a method that improves adaptability without extra supervision. DeLock looks like a promising direction.🔥

Ever post-trained a VLA and watched it ignore every novel instruction? We call this lock-in. Prior fixes bloat datasets with foundation model labels. 🔓DeLock is different: regularized finetuning + contrastive prompts at inference. Result: Pretraining priors preserved.

I always feel frustrated to see the finetuned VLA policy become useless to any other task. We need generalizable, steerable VLA that can perform well on multiple tasks (all the tasks ultimately). Checkout DeLock that elegantly solve this problem!

Recover steering ability of the pre-trained VLA with simple and efficient post training harness👍 Check Suning’s DeLock

New work led by @suning_huang exploring how we can preserve steerability during low-data post-training for VLAs 👀 Details below! ⬇️⬇️

💡n/n Huge thanks to my wonderful collaborators @JiaqiShao0819, @ke_wang123, @QianzhongChen, @JiankaiSun, @GYanjiang and amazing advisors @MacSchwager, @leto__jean for all the ideas, feedback and support that made this project possible!

💡6/n More details here: 🌐 Project website: suninghuang19.github.io/delock_page/ 📄 Paper: arxiv.org/pdf/2604.23121 🎥 Video: youtu.be/erHcKKbIzIA?si…