Coverage: 2026-10-26 00:30 UTC to 2026-10-27 00:30 UTC (2026-10-26 08:30 to 2026-10-27 08:30 Asia/Shanghai). The frozen partial snapshot contains 96 finalized candidates. This edition covers 10 selected items: three priority signals and seven additional briefs. Figures from company statements and paper abstracts remain attributed; this partial finalized-only sample is not an exhaustive industry census.
Also worth knowing
RynnUnify explores a 16B unified embodied model
An arXiv abstract introduces RynnUnify, a 16B-parameter model combining video-language reasoning, physical prediction and grounding, with extensions toward robot control. The authors report a 9.75-point lead over Cosmos 3 Nano on Rynn-VP and a 6.0-point gain for an agentic extension on Rynn-GEU. The world-action extension reportedly succeeds on 51.5% of 20 selected RoboCasa365 tasks and averages 46.7% over three real-robot tasks. These are author-reported benchmark outcomes, not chip-level latency, memory or power measurements.
FLORA proposes an open robotics data-labeling engine
The FLORA abstract describes an open model suite generating outcome rewards, process rewards and timestamped descriptions for robotics data engines and GR00T training. Its authors claim leading results on WGO-Bench, RoboReward and Robometer without giving scores in the supplied abstract. They also claim roughly 125 times lower cost than a frontier-model API alternative and labeling 200,000 hours of data per day using 512 GB300 devices. Cost methodology and independent throughput verification are not provided; these figures should be treated as claims rather than generalizable production benchmarks.
VLA training study flags bf16 updates lost without fp32 master weights
Researchers report that 9 of 12 inspected open VLA codebases use or document a bf16 fine-tuning path without fp32 master weights. In their LIBERO probes, at least 79% of weight updates were rounded away in that setup. With fp32 trainable weights at a lower learning rate, reported LIBERO-Object success changed from 4.3% to 49.3% for SmolVLA and from 5.1% to 80.9% for pi_0.5. These are experiment-specific results, not a blanket claim that every bf16 training recipe fails. The immediate engineering signal is to audit optimizer precision and effective weight updates.
RAPAC-DP addresses delayed execution in diffusion policies
RAPAC-DP conditions a diffusion policy on already committed actions so the next generated action chunk aligns with delayed handover. It uses a lightweight pending-action encoder and LoRA while keeping the base policy frozen. The authors report 450 physical Franka Panda trials, with high-delay mean success improving from 28.9% for RTC to 60.0% for RAPAC-DP. Simulation results also report 81.4% of delay-free performance retained on Kinetix. This is evidence about the tested delay settings, not a universal real-time guarantee.
UCA-Flow reports 3.0 ms action inference
UCA-Flow puts observation, timestep, interval and action tokens into a unified transformer for one-step action generation. The authors report 86.4% average success, a 7.5-percentage-point lead over their strongest baseline, and 3.0 ms task-averaged inference latency. They also report a 43.9-times speedup over DP3 in their evaluation. These figures depend on the tested model, hardware and tasks; the bounded abstract does not establish deployment power, thermal headroom or a specific NPU mapping.
SIREN tests gaps in humanoid safety-filter assumptions
A robotics paper argues that collision guarantees from safety filters depend on correct task- and environment-specific parameters. Its SIREN tool searches for vulnerable goals; the authors report detecting every vulnerable case in the SPARK dataset and identifying three times as many as a baseline. They also report a safety violation on a real humanoid robot, without specifying the hardware or conditions in the abstract. This is a research warning about configuration-sensitive guarantees, not proof that all humanoid safety filters are unsafe.
ViT-H links video generation to humanoid motion feedback
ViT-H generates humanoid reference motions directly from monocular video and uses motion-tracking outcomes for preference-based post-training. Its authors report an autoregressive variant for video teleoperation with latency below 400 ms and experiments on Unitree G1 that outperform reconstruction-and-retargeting pipelines on motion quality and executability. The abstract does not give numerical success rates or a chip-level compute budget. The reported latency should not be equated with a certified control-loop deadline.