Robot policy gains now depend on memory, targeted adaptation, and harder dexterity tests

Several papers give vision-language-action (VLA) policies explicit control over task progress. Harness VLA treats a frozen VLA as a short, retryable contact skill, while a planner handles grounding, transport, staging, and failure recovery. It reaches 82.4% on LIBERO-Pro, compared with 50.0% for the direct frozen baseline. TFP stores an episode-local belief and updates it using elapsed time and interaction events; real-robot object-swap success rises from 3/20 to 15/20. LEEVLA adds task-relevant region weighting and latent future-feature prediction during training, reaching 98.2% on LIBERO without extra inference cost.

Adaptation methods are modifying compact interfaces around pretrained policies. FlowDAgger converts sparse human corrections into latent noise targets for a frozen generative controller. On 12 MetaWorld tasks, mean success reaches 0.78 after 50 rollouts, versus 0.53 for the base policy. Prompt-Driven Exploration rewrites task prompts after inspecting rollout videos; in its microwave case study, canonical-prompt success reaches about 98% while action-noise PPO remains near zero. CLAP places a natural-language action description before numeric controls. Its 2B model reaches 90.8% on LIBERO after one epoch, 14.9 points above the matched baseline.

FabriVLA shows that a 0.89B-parameter model can reach 92.0% episode success on Meta-World MT50 using intermediate visual-language features and gated attention across action tokens. DexVerse exposes a much harder boundary. Across 19 dexterous tasks, the best mean success is 0.34, shared by 3D Diffusion Policy and pi0.5. Every tested method scores zero on PushT, and insertion, knife sliding, and laptop opening remain near zero. Strong results on standard suites therefore provide limited evidence for force control, fine alignment, and multi-stage hand use.