8/8 Beyond Vision: It's still an open question to what extent other sensing modalities are needed and how to best fuse them, but there are promising works in this direction (e.g., Force VLA, Touch-in-the Wild). I’d love to hear what you found most interesting! #NeurIPS #AI

7/8 Safety: VLAs still fail in many situations. Offline safety alignment (e.g., SafeVLA) and combining online failure detection (e.g., FIPER, SAFE) with data-efficient human-in-the-loop learning (e.g., Compliant Residual DAgger, APO) hold a lot of promise to improve VLA safety.

6/8 Inference Speed: Not really a bottleneck anymore for VLAs thanks to new tokenization (e.g., BEAST), caching (e.g., VLA-Cache, EfficientVLA), and fast-sampling sampling methods (e.g., Two-Steps Diffusion Policy, MeanFlow).

5/8 Learning from Videos: This is progressing (e.g., EgoBridge, Object-centric 3D Motion Field, KungfuBot), and it seems to be easier for locomotion than for manipulation, where the embodiment gap is often bigger.

4.5/8 Currently, world models are mainly used in two ways in robotics: Data generation for policy training (BOOM, Newt, RoboScape) and safe policy eval (AutoVLA, ReSim). Incorporating world modeling objectives into VLA training also helps learn better representations (DreamVLA).

4/8 World Models: They have become quite powerful (e.g., MindJourney, PlayerOne) and are starting to become real-time capable. In autonomous driving, they already work well at a larger scale, as showcased by several companies, but things are harder for manipulation due to contact

3/8 RL for VLAs: This is becoming a hot topic. RL for flow matching has been studied a lot recently (ReinFlow, CQN-AS), but sample efficiency remains a challenge. For RL fine-tuning of VLAs on hard real-world tasks, a strong base policy and targeted exploration will be crucial.

2/8 Imitation Learning: Learning from teleoperation data is too inefficient to achieve long-term, general autonomy on its own. This is not particularly surprising, as even for LLMs, SFT (imitation) on all internet data has plateaued, so that many people are now working on RL.

Embodied AI & Robotics research at #NeurIPS2025 - 8 takeaways. 1/8 Continual Learning: The ability to learn from experience is still largely missing in embodied agents. A key problem is that continual deep learning (no forgetting or plasticity loss) does not yet work well.

I’ll be at NeurIPS this week, presenting our work on runtime failure prediction for generative policies. If you’re around and want to chat about research, please reach out! 🕒Friday, 5 December, 4:30 pm 📍Poster 2204 🌐 Website: https://t.co/fD8cH1OKmd #NeurIPS2025 #NeurIPS