Robot policy gains hinge on making scarce experience and action signals usable

Learning from Hindsight relabels failed robot rollouts with the behavior they actually completed, then scores those trajectories against the new instruction. For vision-language-action (VLA) post-training, this keeps 70%–80% of trajectory groups usable, versus 20%–40% with standard methods. On out-of-distribution LIBERO-PRO tasks, it reaches standard training’s final performance in about five steps instead of nearly 30. With 160 physical-robot rollouts, success reaches 56%, compared with 22% for standard group-relative policy optimization.

CD-LAM targets action-conditioned world models whose learned action codes also respond to camera motion, backgrounds, or untouched objects. Embodiment-weighted reconstruction, contrastive training, and zero-transition calibration reduce these shortcuts. Camera-shift responses fall from 0.555 to 0.156 horizontally and from 0.545 to 0.110 vertically. The 14-billion-parameter model matches the DreamDojo reference with more than 12 times fewer robot-action adaptation updates.

Two papers treat execution quality as a property of the action representation and training objective. B-spline Policy predicts continuous action curves that can be resampled at higher control rates. In table cleaning, it cuts average completion time from 23.57 to 11.80 seconds while success rises from 13/20 to 14/20; aggressive 4× speed causes failure on one stacking task, exposing controller limits. PAC-ACT aligns reinforcement learning with eight-step action chunks and constrains updates near the pretrained policy. On a precision-contact task, it reduces the share of force readings above 60 N by 46 times.

TactiDex records 5.1 million synchronized frames of whole-hand pressure, hand motion, object pose, task phase, and text across 757 demonstrations. Its residual policy uses human pressure patterns to guide contact timing and force regulation, though the available evidence does not provide numerical success rates. For aerial perception, ATRNet-LUDO supplies 121,000 panoramic images and 1.21 million target patches across 40 outdoor scenarios. Active UAV observation improves recognition by about 20 percentage points over passive perception, while its predictive world model adds 2–3 points over deep reinforcement-learning baselines at similar motion cost.