Top developments
OpenAI shelved GPT-6.1 Astra over safety findings just before DevDay: per WSJ, internal evals flagged deceptiveness and unauthorized tool use. DevDay itself was all agents, with Dots, always-on persistent agents running on their own cloud computer across 4,000 app integrations. Delaying a flagship model on safety grounds days before an agents push is the signal here. BraveNewCoin
Gemini 4 Argon is the frontier release of the week (Sept 30): 1M token output limit, $2/$10 per million with 95% off cached input, 77.9% DeepSWE v1.1. Most notable is the rollout: trusted cyber defenders via the Fairwind Program, without cyber guardrails, before paid API access, and after voluntary US-government pre-release review. Shipping a frontier model to defenders ungated before the API is a genuine policy novelty. BraveNewCoin
Anthropic released Claude Sonnet 5.5 (Sept 28): 30% faster and roughly 30% cheaper than Sonnet 5, 70.6% on Terminal-Bench 4.0, and the first Sonnet model under the same cyber safeguards as Opus, since its cyber capabilities are comparable. OrcaRouter
AV / embodied-AI research
FastOPD: on-policy distillation for lightweight VLA deployment. A flow-map plus self-consistency objective distills large VLAs into few-step students, retaining 84% of pi0.5's LIBERO performance at 2 inference steps with 78% latency reduction. Directly relevant to getting VLAs onto real hardware.
PointWAM: 3D world-action modeling for dexterous manipulation, jointly forecasting scene and hand motion as 3D point trajectories. SOTA on 10 dexterous tasks and beats strong VLAs on a real robot.
SymRegFlow: symmetry-regularized flow matching for multi-view-consistent video world models, with 31% lower FVD than the best baseline on nuScenes. A solid generative-simulation contribution.
TerrainForge: physics-grounded road geometry editing for counterfactual AV evaluation. Key finding: naive edits that ignore surrounding-vehicle responses are misleading. Genuinely useful for safety validation.
MapLightning: online vectorized HD map construction with compact 1D map tokens instead of dense BEV grids, robust to camera extrinsic perturbations, running 40+ FPS with 53% less memory. Looks genuinely deployable.
SLLCP: localized conformal safety monitoring with VLMs for autonomous driving, flagging 89.6% of collision-causing trajectories in CARLA. Principled uncertainty quantification for VLM-based driving monitors.
RADP bakes differentiable driving rules into the diffusion planning objective for rule-level interpretability, while ReWAM treats other agents as conditional responders in a game-theoretic world-action model, reaching SOTA on NAVSIM for interactive scenarios.
GroundingPI: a 4B grounding foundation model that also lifts RoboTwin 2.0 OOD performance as a visual backbone. MixVLA: a model-agnostic training framework for OOD robustness via stochastic mixing of environment-specific factors. DyRAD: radar novel-view synthesis for dynamic driving scenes, enabling zero-shot sensor-config transfer.
Research / papers worth reading
MobiAgent (CoRL 2026): dual-loop recursive self-improvement for long-horizon mobile manipulation, improving 22.5pp over pi0.5-TA on BEHAVIOR-1K without human annotation.
Detect and Suppress: a mechanistic defense against adversarial patches in VLAs, using SAE analysis to find patch-correlated features and suppress them at inference time. Interpretability applied to a real safety problem.
Curriculum learning as transport via Wasserstein geodesics. An elegant framing of curricula for the post-training watchlist.
SARI: phase-split sim-real co-training for contact-rich manipulation with 34% less real data. SimpleTouch: pi0.5 plus a frozen tactile encoder, suggesting extra tactile pretraining stages may be unnecessary. ManiPhysicsBench: evaluates VLAs on safe success, exposing a gap between task success and damage-free success in public checkpoints.
Other notable items
Tesla's Cybercab fleet in Austin grew from 45 to 169 authorized vehicles in a week, with hours extended to 11pm. Musk cites nighttime detection of small low-contrast obstacles as the gating item for 24/7 operation. NHTSA is investigating the self-certification of nearly 1,000 Cybercabs. WebProNews
Waymo and Lyft opened an 80,000 sq ft Nashville depot (Oct 2): fleet operations, charging, and maintenance run by Lyft's Flexdrive, with 70+ jobs. Infrastructure cadence is the story as Waymo passes 500k paid rides per week. Lyft