Embodied AI work is concentrating on the action loop

Safety work in vision-language-action models is getting more concrete. VLA-Forget edits the visual encoder, cross-modal projector, and upper action layers in stages, then uses retain, forget, mismatch, and feature-preservation losses to remove unwanted behavior without wiping out task skill. The reported numbers are strong for this kind of intervention: on OpenVLA-7B with Open X-Embodiment it posts FC 93, RC 91, and TSR 78, with lower safety-violation recovery after quantization than NPO and much better retention than GA. That matters because the failure target is a robot action, not just a text output.

Inference-time control is a live optimization target. AAC changes action chunk size on the fly by reading entropy over sampled candidate trajectories, then replans sooner when the model is uncertain and executes longer chunks when predictions stay stable. The gains are modest but broad: GR00T rises from 59.7% to 62.0% on RoboCasa, from 94.1% to 95.0% on LIBERO, and from 88.8% to 92.8% on LIBERO-Long. The pattern matches the last few days of robotics work in this corpus: progress is coming from tighter control over the action interface and replanning loop.

Joint video-and-action generation is still one of the clearest routes to stronger driving planners. DriveVA predicts future video latents and ego trajectory tokens in one generative process, and the paper ties a large part of its gain to dense video supervision. On NAVSIM v1 it reaches 90.9 PDMS, above DiffusionDrive at 88.1, and it reports large zero-shot transfer gains on nuScenes and Bench2Drive/CARLA v2, including an 83.3% collision-rate reduction on nuScenes against the stated world-model baseline. The notable point is practical: the model claims near-optimal closed-loop performance with only two sampling steps.

Dexterous manipulation work is also leaning on sparse structure that humans can inspect. GRIT uses a 30-way grasp taxonomy plus wrist orientation as the high-level command, chosen by a vision-language model from the scene and task, then executes with a taxonomy-conditioned control policy. It reports 87.9% overall success and tests on 373 novel objects after training on 30 YCB objects. The paper also emphasizes user control: when a grasp strategy is wrong, the interface gives a concrete handle for changing it.