Embodied AI research this week treats robot policies as systems that must survive real control conditions. Vision-language-action (VLA) models are tested through visual corruptions, 3D contact cues, memory, latency, and…
VLA teams can add small physical regression benches, target-state logging, and RGB-D geometry paths to check whether manipulation policies still work under real control conditions.
April 22’s research is strongest where coding work meets concrete checks from real use, project scaffolding, and direct execution. SWE-chat grounds coding-agent claims in kept code and user pushback.
Coding-agent evaluation is moving closer to what teams can verify in their own repositories and pipelines. The most usable directions here are a commit-linked scorecard for kept code and review friction, a narrow…