This day’s robotics set is easy to read: evaluation gets stricter, and model design follows the same pressure. HazardArena measures whether a VLA can tell a feasible action from a dangerous one.
Robotics work in this window points to three concrete workflow changes: evaluate VLA safety with stage-wise hazard progression and a refusal gate, add future tactile latent prediction to humanoid behavior cloning on…
April 5 is a strong embodied-AI day centered on action control. The clearest papers improve what happens between perception and execution: safer VLA policies, adaptive action chunking, and joint video-action planning…
Embodied AI work here is getting more concrete at the control interface. The clearest near-term changes are a post-training unlearning workflow for deployed VLA policies, an inference-time adaptive chunking layer for…