Coverage: 2026-10-30 00:30 UTC to 2026-10-31 00:30 UTC (2026-10-30 08:30 to 2026-10-31 08:30 Asia/Shanghai). The frozen partial, finalized-only snapshot contains 153 candidates; this edition covers 10 selected or near-selected articles (3 priorities and 7 briefs). Most research items rely on supplied abstracts; reported results, company claims and future plans are attributed, not independently verified.
Also worth knowing
HERON reports INT8 reachability mapping on Jetson AGX Orin
HERON uses a 3D U-Net trained from cuRobo reachability maps and runs with TensorRT INT8 on Jetson AGX Orin. Its authors report a 2.56 m cubic workspace at 2 cm resolution, averaging 115 ms per map and a system rate of 8.57 Hz, versus about 90 minutes for the referenced cuRobo computation. They report 95% voxel agreement within one of 51 classes for FP32 and within three for INT8; boundary regions degrade most. These are paper-reported results on one platform, with power, model footprint and portability undisclosed.
Ego-OSCAR releases a low-BOM synchronized capture rig
Ego-OSCAR describes open hardware with a component-and-print BOM below USD 200, pairing synchronized global-shutter stereo cameras, a six-axis IMU, an embedded Linux encoding board and a real-time MCU. The authors report a 700-microsecond camera-inertial clock bridge residual and a dataset of roughly 550 hours across 1,462 sessions and 25 contributors. CAD, firmware and synchronization tools are released. Predicted 3D hand reconstructions are not human-verified ground truth; the USD 200 figure is BOM, not a selling price.
UMI-Live links mobile bimanual capture to VLA training
The UMI-Live authors present an open-source bimanual data-collection system integrating cameras, LiDAR-inertial sensing, onboard compute and power, without external tracking at collection sites. An asynchronous pipeline reconstructs SLAM data and publishes training-ready sets; software supports dual-YAM teleoperation and VLA inference. They report 40 hours of data across three tasks and bimanual pick-and-place success of 10/10 for pi_0.5 and 8/10 for MolmoAct2 after fine-tuning. The abstract does not disclose SoC model, power, sensor rates or repeatability beyond those tests.
VLA-Corrector proposes conditional replanning of action chunks
VLA-Corrector compares predicted and observed visual features with a lightweight latent-space monitor. Persistent mismatch shortens an action chunk and triggers corrective replanning using Online Gradient Guidance; otherwise longer chunks continue. Its authors say the method can integrate with different VLA backbones without retraining their weights. The abstract claims better robustness while retaining chunk efficiency, but provides no numeric success gain, monitor overhead, latency or hardware benchmark. It is a research proposal, not a measured edge deployment.
WorldFlow jointly predicts future video and 3D motion
WorldFlow uses a Mixture-of-Transformers backbone to generate video and aligned 3D displacement trajectories. The authors report training on 1.44 million object-level samples and a WorldFlowBench evaluation set of 1,678 samples from 11 sources. On SpatialCog they report 0.060 m average displacement error, 54.2% below the strongest external baseline; against a flow-only variant, reported ADE and FDE reductions are 55.9% and 68.5%. These are abstract-reported benchmark outcomes, not demonstrated robot-task gains or edge SoC throughput.
SKiD-SLAM field-tests distributed mapping under a radio budget
SKiD-SLAM combines inter-robot place recognition, relative poses and distributed trajectory fusion. The researchers describe validation on public data and a rover simulator, then field mapping with wheeled rovers in a Jeju lava-tube cave and quadrupeds at a Yeoncheon crater-and-rock site. They report operation within a 250 kbps inter-robot communication budget. This is field-research evidence, not a production deployment; localization accuracy, onboard compute load and fleet size are not quantified in the supplied abstract.
Percepto shifts inspection analytics toward hardware-agnostic SaaS
The Robot Report says Percepto launched Axion, a SaaS platform that organizes inspection data by asset and can ingest supported third-party drones, mobile or fixed sensors and historical records. Percepto claims more than 160 AI models and natural-language inspection queries. It plans to keep selling and supporting drones and docks, while planning US manufacturing. Target use cases include oil, gas and power infrastructure. Pricing, customers, model accuracy and edge-versus-cloud inference split are undisclosed; the model count is a company claim.