Robot policy progress depends on memory, reusable failures, and executable action control

Policies are being given explicit state about task progress. TFP maintains a continuous-time latent belief and raises LIBERO Long-10 success from 92.4% to 97.0%; on a real object-swap task, success rises from 3/20 to 15/20. Harness VLA places a memory-guided planner around a frozen controller, using stored traces, re-grounding, staging, and retries. It reaches 82.4% on LIBERO-Pro, compared with 50.0% for the direct frozen baseline. LEEVLA adds task-relevant region weighting and latent future-feature prediction during training, reaching 98.2% average success on LIBERO without extra inference-time memory or computation.

Collected experience is being reused more aggressively. Learning from Hindsight relabels failed rollouts with the behavior the robot actually completed. On out-of-distribution LIBERO-Pro tasks, it reaches standard reinforcement-learning performance in about 5 training steps versus nearly 30, and records 56% real-robot success after 160 rollouts versus 22% for the baseline. CD-LAM removes backgrounds, camera motion, and unrelated objects from latent action codes learned from video. It matches its reference with more than 12 times fewer robot-action adaptation updates, making unlabeled video a cleaner source of action supervision.

Action representations are being evaluated as control interfaces with measurable timing and safety limits. B-spline Policy predicts continuous curves that can run at different speeds; on long-horizon table cleaning, it cuts average completion time from 23.57 to 11.80 seconds while success changes from 13/20 to 14/20. PAC-ACT aligns reinforcement learning with eight-step action chunks and reduces force readings above 60 N by 46 times on its Contour task. DexVerse exposes the remaining ceiling: across 19 tasks, the top mean success rate is 0.34, and all four tested policies score zero on PushT.