New robotics papers with experiments on physical robots: manipulation, locomotion and robot learning.
This issue has twelve items. Three are imitation-learning papers on keeping action-chunk replanning consistent and fast, and two look at how VLAs are post-trained and what they encode. Most worth reading: the probing study showing VLAs encode physical properties weakly, and the finding that augmentation should touch only the critic in VLA RL post-training. Most items do not mention released code or data, and they are flagged below.
InterMimicGen retargets human-object interaction mocap data to humanoids with dexterous hands and trains a generalist tracker. A data flywheel then adds task-preserving edits and keeps only the variants that succeed in simulation. The paper reports contact-preserving retargeting, broad tracking with a single policy, and transfer to real robots. No code or data release mentioned.
RV-ICL turns one demonstration video per task into a hierarchy (keyframes, phases, moments, clips) that an LLM agent orchestrating frozen VLA policies navigates through read-only tools. Built on RPent, it raises success from 92.6% to 96.5% on LIBERO-PRO and from 86.7% to 95.8% on LIBERO-Plus. Training-free. No code or data release mentioned.
Combines a physics-residual cRVAE cloth model with an MPPI and MPC planner for a dual-arm mobile manipulator carrying cloth with a human. The reduced-order model tracks about as well as the full 17-DoF model with roughly 80% less computation. Whole-body refinement cuts final cloth deformation from 0.93 m to 0.28 m, and the baseline collides where this keeps clearance. No code or data release mentioned.
Streams that already watch this field. Follow one as it is — it costs nothing extra.
Humanoid and general-purpose robot companies: launches, funding rounds and first real deployments.