Coverage: 2026-10-27 00:30 UTC to 2026-10-28 00:30 UTC (2026-10-27 08:30 to 2026-10-28 08:30 Asia/Shanghai). Frozen partial snapshot: 183 finalized candidates; 10 selected articles covered, including 3 highlights and 7 briefs. Research metrics are author-reported, and commercial claims are attributed to their sources. This is a curated, partial-coverage edition, not an exhaustive survey.
Also worth knowing
Embotech and BMW report factory-scale automated vehicle movement
The Robot Report says Embotech and BMW reached 1 million automated vehicle movements at four sites in Germany and the U.K., with about 3,500 movements daily and availability above 99.5%. An Embotech executive describes infrastructure-mounted Hesai lidar and offboard perception, with servers optimizing trajectories 10 times per second. The company also describes a separate deterministic safety boundary. These are reported operational figures, not an independent audit.
Luxonis announces OAK 4 FFC vision computer
An Open Robotics Discourse announcement describes the OAK 4 FFC as a 70 mm × 53 mm vision computer supporting up to five cameras, synchronization, recalibration, depth and AI. A developer kit is advertised at $849 USD with shipping planned for November, and an open-source reference design is linked. Luxonis also upgraded its OSRA membership to Silver. The kit price is not a production-module or SoC price; camera throughput, chip identity and power were not disclosed.
Memory-aware map loading keeps dense maps and language models together
Researchers propose tracking each keyframe's resident bytes and loading map data by semantic room or corridor zones, rather than fixed frame counts. In tests on two Jetson boards, authors say memory budgets held and language-model throughput fell by at most 0.8%; without the policy, memory exhaustion could terminate a model or map. They also report localizing 6–9 percentage points more query frames than a matched geometric partition. Absolute memory and power figures are not given.
RynnVLA-Latent scales robot learning from unlabeled interaction video
RynnVLA-Latent learns 3D-grounded latent actions from video transitions without native robot control labels. Authors report training on 30.1K hours of human, robot and simulated video, producing a 4B-parameter policy. They claim 98.8% on LIBERO and 84.7% on LIBERO-Plus. These are paper-reported benchmark results, not proof of edge deployability; model precision, onboard latency and power are undisclosed.
OSWAM reports one-step world-action generation
OSWAM uses structured imagination tokens and future-conditioned distillation to reduce repeated action denoising. Authors report approximately 3x speedup over direct-policy world-action-model baselines and compiled inference latency as low as 40 ms on an NVIDIA RTX 5090. The RTX 5090 measurement is a desktop GPU result, not an edge SoC benchmark; action quality and portability need further validation.
ForceDelta-VLA distills force corrections instead of regenerating actions
ForceDelta-VLA trains a lightweight correction policy from force-conditioned and force-agnostic teacher predictions, adjusting reference actions with recent force history. Across nine contact-rich tasks, authors report 82.2% mean success, compared with 54.4% for ForceVLA and 70.6% for a direct temporal teacher; mean peak contact force was approximately 26% lower than ForceVLA on successful trials. Update frequency, onboard latency and deployment costs are not disclosed.
Xpanner announces autonomous pile driving for solar installations
Xpanner announced the X1 Piling R6 for utility-scale solar and battery-storage construction. According to The Robot Report, the system combines GNSS, onboard sensing and an automation controller, while an operator supervises and can take remote control. A web dashboard logs pile placement and machine status. The company describes purchase or rental plus monthly software licensing; independent productivity figures, customer orders and controller SoC details were not disclosed.