Embodied control papers are fixing concrete bottlenecks in action, planning, and transfer

Robot control papers are getting more explicit about where action quality is lost and how to recover it. The clearest evidence comes from two directions. One paper shows that better visual encoders do not reliably help a vision-language-action policy when actions are squeezed through discrete tokens. On LIBERO-10, Diffusion Policy climbed from 36.4% to 57.6% with a ResNet-18 to SigLIP upgrade at size M, while OAT moved from 53.8% to 57.4%. Another paper improves long-horizon world-model control by splitting planning into latent macro-actions and low-level actions. HWM raised real Franka pick-and-place success from 0% to 70% and drawer tasks from 30% to 70%, with lower planning cost claims in the same report.

Low-data manipulation work is leaning on predictive video structure, not just static image features. MV-VDP predicts future multi-view RGB videos and action heatmaps together, then lifts the heatmaps back into 3D action estimates. In Meta-World with 5 demos per task, it reached 89.1% average success, ahead of Track2Act at 67.4% and DreamZero at 61.1%. The real-robot table is smaller and mixed, but still concrete: 10/10 on Put Lion, 7/10 on Scoop Tortilla, and 4/10 on Push-T, with weaker numbers for the listed baselines. The pattern is that scene dynamics and 3D consistency are being trained directly into the policy output.

Sim-to-real papers in this period focus on preserving control semantics at deployment time. For quadrupeds, DreamTIP adds task-invariant properties such as contact stability and terrain clearance to a Dreamer world model. It reports a 28.1% average gain across 8 simulated transfer tasks and strong real-world gains on Unitree Go2, including Climb 52 cm from 10% for WMP to 100%. For driving, Sim2Real-AD breaks transfer into observation bridging and action remapping. Its zero-shot real-vehicle results on a Ford E-Transit are 90% for car-following, 80% for obstacle avoidance, and 75% for stop-sign interaction, without real-world RL training data in the reported setup.

VLA efficiency work is getting more surgical about when a large model needs to run. SV-VLA keeps a heavy planner for action chunks and adds a small verifier that checks each step against the current observation. On LIBERO, it reports an average gain from 79.5% to 90.90% over the open-loop baseline across three subtasks. The evidence is narrower than the robotics planning papers because the excerpt does not include per-task scores or latency tables, but the method is clear: chunk first, verify online, and replan only when the current state disagrees with the planned action.