This edition uses the finalized-only partial snapshot for 2026-10-08 00:30 UTC to 2026-10-09 00:30 UTC (2026-10-08 08:30 to 2026-10-09 08:30 Asia/Shanghai), as of 2026-10-09 00:40:46.098 UTC. The snapshot contains 156 finalized candidates; this editorial selects 11 supported articles. The feed is marked partial, so this is not a comprehensive market census. Most research findings are abstract-level author claims; commercial and developer claims are attributed to their reports or announcements.
Also worth knowing
VPP2 links stronger video prediction with action performance
The VPP2 arXiv abstract describes event-level video pretraining, a single-step fixed-horizon visual planner and a Mixture-of-Transformers action module. Authors report VPP2-14B beating Cosmos3-64B by 11.0 percentage points on open-ended video instruction following and improving real-world zero-shot ALOHA manipulation success by 18.5 points over their strongest baseline. Benchmark-specific post-training results are also claimed; independent replication is not established.
RealtimeWAM accelerates inference without retraining
The RealtimeWAM arXiv abstract describes parallel execution, caching and selective Transformer refinement. On RTX 4090, FastWAM and OpenWAM are reported at 24.09 ms and 63.09 ms mean latency, respectively, corresponding to 8.90x and 10.67x speedups. Their average success rates are 82.75% and 87.41%, within 0.02 and 0.53 percentage points of native inference; the authors also report real-task gains. No equivalent edge-SoC timing is given.
RoboJEPA tests scaling laws for latent robot world models
The RoboJEPA arXiv abstract reports a JEPA-based world model trained across 12 robot embodiments, including an 8-billion-parameter predictor. Authors describe a second-order power law for latent rollout imagination error and a correlation with downstream planning, plus zero-shot long-horizon planning on real hardware from a single goal image. They state checkpoints and code are released; the bounded evidence does not independently verify release assets or provide edge inference costs.
RoboQuest finds information-gathering remains a bottleneck
The RoboQuest arXiv abstract introduces 10 physical search, inspection and testing tasks. Across 5 frontier multimodal agents and a fine-tuned pi_0.5 policy, the best agent succeeds in 23% of episodes and the fine-tuned policy almost never succeeds, according to the authors. Their analysis says agents often stop before collecting decisive evidence, even when isolated execution skills are available. This is a benchmark diagnosis, not a field reliability measurement.
SAFE demonstrates acoustic grasp monitoring on Jetson Orin Nano
The SAFE arXiv abstract combines 2 passive PVDF acoustic sensors and motor feedback on a Fin Ray gripper, without vision or prior material knowledge. A 79-dimensional feature classifier reports Alert-F1 of 0.884. A grasp controller running at 104 Hz on Jetson Orin Nano reaches 91.3% success in 46 closed-loop trials; novel-object success is 82.4%. This is a concrete embedded prototype result, not a product power or cost specification.
Jabil flags manufacturing and compute constraints in humanoid scaling
The Robot Report says Jabil is working with Apptronik on Apollo manufacturing deployment. A Jabil executive describes customer robot deployments as low-volume today and tens-of-thousands scale as a future phase, not a shipment disclosure. He highlights edge-compute demand, Jetson adoption, compute and memory costs, wiring robustness, modular power and unresolved humanoid safety certification. These are attributed industry observations and forecasts.
Helm.ai reports $70 million in signed model contracts
The Robot Report attributes to Helm.ai a claim of $70 million in signed commercial foundation-model contracts over 12 months, across automotive OEMs, Tier 1 suppliers and industrial automation. Its CEO calls automotive the flagship, with mining and construction projects and a path toward break-even. The article describes SAE L2–L4 programs and multimodal camera, lidar and radar work. Signed contracts are not equivalent to recognized revenue or independently verified deployments.
Uranus announces a data-driven robot simulation model and upcoming technical talk
An Open Robotics Discourse announcement describes Uranus as a joint-trajectory-conditioned autoregressive diffusion world model using multi-view robot observations. The authors claim more than 3,300 hours of robot data, a 1.3B-parameter model and 24 FPS serving. A D-Robotics Frontier technical talk is announced for 2026-10-15 at 20:00 GMT+8. Code, weights and data are intended for open source release, but the post does not establish that they are already available.
Watching next
Watch whether reported WAM and flow-policy speedups survive matched on-device tests with explicit latency, memory, power and task-success measurements; today's results span different models and hardware.
Watch whether toolchain integration and manufacturing plans yield measured robot deployments, repeatable validation and unit economics rather than announcements or forward-looking production targets.