Coverage: 2026-10-28 00:30 UTC to 2026-10-29 00:30 UTC (2026-10-28 08:30 to 2026-10-29 08:30 Asia/Shanghai); frozen as_of 2026-10-29 00:40:31.778 UTC (2026-10-29 08:40:31.778 Asia/Shanghai). This partial finalized-only snapshot contains 124 candidates, of which 10 selected or near-selected articles are covered: 3 highlights and 7 briefs. Most research claims are from abstracts or developer reports rather than independently reproduced measurements; coverage is not a comprehensive survey.
Also worth knowing
PDRVLA conditions robot actions on predicted physical dynamics
An arXiv abstract describes physical tokens derived from vision-language features, supervised by future RGB, depth and optical-flow features without online video generation. Its authors report simulation success of 99.3% on LIBERO, 89.9% on LIBERO-Plus, 93.6% on RoboTwin 2.0 and 81.5% on SimplerEnv. These are reported benchmark results, not verified robot deployment or edge latency measurements.
VLA-Stem adds a safety correction layer without retraining the base VLA
VLA-Stem converts proposed action chunks into motion references and adjusts execution using proprioception and local obstacle information. Its authors report 52.8% collision-free task success across six constrained simulation settings versus 23.7% for the strongest compared safety baseline. They also report fewer collisions with static and moving obstacles on real robots, without numerical real-world reduction figures in the abstract. This is a research safety mechanism, not a safety certification.
RAPID tests perception and tool orchestration in EV battery disassembly
The RAPID research platform combines an industrial gantry robot, RGB-D sensing and a nut-running tool. In 204 fastener-removal evaluations, the authors report 97% success for taught poses, 57% for one-shot vision and 83% for visual servoing. Their structured tool interfaces reportedly completed 100% of tested tasks, while automatic ROS service discovery had a 43.3% failure rate. The results concern a research setup, not validated factory production.
Small-robot study exposes the cost of limited demonstrations
An arXiv study compares ACT, Diffusion Policy, SmolVLA and π0.5 on a physical SO-101 arm across 12–200 demonstrations and 1,340 evaluation episodes. ACT reportedly rose from 1/40 successes with 12 demonstrations to 37/40 with 100; π0.5 action-expert fine-tuning scored 16/40 with 12 and 28/40 with 25. The authors caution that fixed training length confounds dataset size with repeat exposure, limiting causal conclusions.
FAWAM introduces force feedback into world-action models
The updated FAWAM arXiv abstract describes conditioning action generation on historical 6-axis force and torque signals, while predicting future actions and end-effector wrench for online correction. Authors report average real-world contact-task success improvements of 36.25% versus vision-only baselines and 21.25% versus prior force-aware baselines. The abstract does not clarify whether these are relative gains or percentage-point changes, and supplies no edge inference measurements.
Zenoh and RaptorQ prototype targets degraded robot links
An Open Robotics Discourse developer describes pairing Zenoh with RaptorQ for outdoor robots on non-line-of-sight links with reported packet loss of 15–30%. The prototype separates bulk assets from signed commands, with a default 128 kbps pacing rate for low-priority data. A local demo at simulated 25% loss reportedly recovered all blocks without retransmissions. This is developer-supplied demonstration evidence, not a field reliability benchmark.
ASPEN reduces repeated navigation VLA inference through command reuse
ASPEN predicts whether an existing navigation command can persist, reusing it at eligible boundaries rather than always calling the frozen base VLA. The authors report 9.5% fewer VLA calls and 8.1% lower estimated inference time including scheduling overhead. On RxR-CE val-unseen, 92.1% of episodes retained identical baseline action sequences. These are research evaluation claims; hardware power savings and deployment latency are not established.