Robotics papers make execution supervision concrete

Long-horizon robot work centered on explicit recovery logic inside the control loop. LoHo-Manip splits planning and execution, then replans from the current image with a remaining-plan memory and a 2D trace prompt. That design produced 63.1 on RoboVQA and 56.7 on EgoPlan2, ahead of named VLA baselines. ReCAPA attacks the same failure mode from another angle. It predicts mismatch across actions, subgoals, and full trajectories, and reports 58.65 on VisualAgentBench plus AI2-THOR success rate 0.75. The common point is simple: papers are putting progress checks and correction signals into execution, not leaving them outside the policy.

World models are getting used as training workspaces, not only predictors. Hi-WM rolls a policy forward in a learned simulator, lets a human intervene near failure, and adds those corrections back into post-training. On three real robot tasks, average success improved by 37.9 points over the base policy and 19.0 over a world-model closed-loop baseline. The paper also reports Pearson r = 0.953 between world-model evaluation and real performance. That makes the simulator useful for deciding when intervention data is worth collecting, not just for offline scoring.

Several papers made physical structure more explicit at the action level. CorridorVLA predicts a few future end-effector anchors and constrains generated actions to stay inside a spatial corridor. That raises SmolVLA from 86.5% to 90.95% on LIBERO and from 45.37% to 57.74% on LIBERO-Plus, with similar gains for GR00T. FingerViP adds five fingertip cameras to a dexterous hand so the policy can see contact regions that wrist or third-person views miss. It reports 80.8% overall success across four real-world tasks, including confined and occluded settings. This day’s robotics papers kept adding structure close to contact: spatial anchors, fingertip views, and signals tied to the actual motion geometry.

Evaluation pressure also tightened. The BEHAVIOR1K audit argues that final-state metrics hide unsafe execution and unstable reproduction. Across 500 reviewed runs, grasp failure was the most common error and collisions were frequent. When safety penalties are added, average RLC score drops from Q = 0.256 to sQ = 0.239, and Comet drops from 0.192 to 0.173. A separate manipulation paper, Wiggle and Go!, fits the same concern from the task side: it uses one low-risk probing motion to estimate rope dynamics before acting, then reaches 3.55 cm real target accuracy against 15.34 cm without parameter-informed control. The message is practical. Better robotics results now depend on safer evaluation and safer information gathering before commitment.