🧠 NVIDIA AVO Scores 100% on ARC-AGI-3, Setting a New Bar for Long-Horizon Autonomous Agents

• ●NVIDIA’s AVO agent achieved a perfect 100% on ARC-AGI-3 — the benchmark designed to resist pattern memorization — marking the first time a general-purpose architecture has saturated this frontier test.
• ●AVO’s architecture is built for long-horizon autonomy, meaning it maintains coherent goal-directed behavior across extended task sequences rather than single-step inference, a critical gap most current agents fail to close.
• ●With ARC-AGI-3 now saturated, the field will need harder benchmarks fast; watch for NVIDIA to position AVO as the backbone for physical AI deployments where multi-step reasoning meets real-world action.
• ●🔎 Read More →
• What matters: A perfect ARC-AGI-3 score signals that general-purpose agent architectures are no longer a research aspiration — they’re a shipping product.

🧪 Technology, Research & Innovation

🧠 AAMAS 2026 Best Blue Sky Paper: Foundation World Models That Let Agents Learn, Verify, and Adapt Beyond Static Environments

• ●Florent Delgrange’s award-winning paper proposes foundation world models that allow agents to reliably adapt when their environment changes — directly attacking the brittleness that plagues current sim-trained systems.
• ●The key technical contribution is pairing learned world models with formal verification, so agents can check whether their internal model still matches reality before committing to an action plan.
• ●If this framework scales, it could replace the current sim-to-real transfer bottleneck with agents that self-certify their own situational awareness — a prerequisite for safe deployment in unstructured environments.
• ●🔎 Read More →
• What matters: Verification-aware world models could be the missing link between impressive lab demos and agents that stay reliable when the real world stops cooperating.

🚗 Survey: LLMs Are Reshaping Autonomous Driving Across Perception, Planning, and Human-Vehicle Interaction

• ●A new survey catalogues how large language models are expanding AV capabilities beyond classical pipelines — covering perception, scene understanding, motion planning, and natural-language interfaces with passengers.
• ●The technical finding that stands out: LLMs enable open-vocabulary scene reasoning, letting AVs handle edge cases that rule-based and closed-vocabulary systems simply cannot classify or respond to.
• ●The survey frames LLMs not as a replacement for sensor stacks but as a reasoning layer on top — expect hybrid architectures combining LLM planners with deterministic safety modules to dominate near-term AV design.
• ●🔎 Read More →
• What matters: LLMs are becoming the reasoning backbone of autonomous driving, turning long-tail edge cases from system failures into solvable inference problems.

🚀 Product, Hardware & Model Launches

🚗 Waymo Reveals Its First Custom Chip: 5nm TSMC Automotive Silicon Now Deployed in Every Robotaxi

• ●Waymo disclosed its first in-house custom chip, built on TSMC’s 5nm automotive process node and already installed across its entire robotaxi fleet — a significant vertical integration move the company had not previously publicized.
• ●Designing at 5nm for automotive-grade reliability means Waymo is optimizing compute density and power envelope specifically for its sensor fusion and planning workloads, rather than adapting general-purpose silicon.
• ●Custom silicon gives Waymo a hardware moat: proprietary chips mean faster iteration cycles, lower per-unit compute costs at scale, and reduced dependency on third-party suppliers as the fleet expands.
• ●🔎 Read More →
• What matters: Waymo joining the custom-silicon club signals that AV leaders now see compute architecture as a core competitive differentiator, not a commodity input.

🤖 Schaeffler to Mass-Produce Strain Wave Gearboxes for Humanoid Robots Starting 2027

• ●Schaeffler announced it will begin mass production of strain wave gearboxes for humanoid robots in 2027, using a forming manufacturing process specifically engineered to hit the volumes the humanoid market is projecting.
• ●Strain wave gearboxes are the preferred actuator component for humanoid joints because they deliver high torque-to-weight ratio and near-zero backlash — but they’ve historically been expensive and slow to produce at scale.
• ●Schaeffler’s forming-based process is designed to break that cost curve; if it delivers, it removes one of the last major supply-chain bottlenecks standing between humanoid prototypes and commercial-scale deployment.
• ●🔎 Read More →
• What matters: Mass-produced strain wave gearboxes in 2027 could do for humanoid actuators what commodity NAND did for storage — collapse costs and unlock volume.

💰 Business, Startups & Investment

🤖 LG and NVIDIA Partner to Build a Humanoid Robot, Combining LG’s Manufacturing Scale with NVIDIA’s AI Stack

• ●LG and NVIDIA have announced a partnership to co-develop a humanoid robot, pairing LG’s consumer electronics and manufacturing infrastructure with NVIDIA’s robotics AI platform.
• ●The combination is technically significant: NVIDIA brings its Isaac and GR00T-based AI stack while LG contributes proven high-volume hardware production — exactly the pairing the humanoid field has lacked.
• ●This is LG’s clearest signal yet that it is repositioning from appliance maker to robotics OEM; watch for the partnership to accelerate NVIDIA’s strategy of becoming the de facto AI OS for humanoid platforms.
• ●🔎 Read More →
• What matters: When a consumer electronics giant bets its next platform on humanoids, it’s no longer a niche robotics story — it’s a manufacturing industry pivot.

🧠 Ouster Tells Investors Physical AI Has Crossed the Line from Demo to Deployment

• ●NVIDIA partner and lidar maker Ouster stated exclusively that physical AI is now transitioning from demonstration to active deployment — a market-stage call that carries weight given Ouster’s sensor-level visibility into real-world robot programs.
• ●For Ouster, this shift means lidar demand is moving from pilot-program quantities to production-volume purchase orders — a fundamentally different revenue profile and a leading indicator for the broader physical AI supply chain.
• ●Investors should treat this as a demand-signal data point: when a sensor supplier sees demo-to-deployment inflection, the hardware procurement cycle that follows typically runs 12–18 months ahead of public product announcements.
• ●🔎 Read More →
• What matters: A sensor supplier calling the demo-to-deployment inflection is one of the clearest early signals that physical AI capex is about to scale.

⚡Agent Benchmarks::NVIDIA AVO’s perfect ARC-AGI-3 score means the field’s hardest general-intelligence benchmark is already obsolete — harder evals are urgently needed.

⚡AV Silicon Strategy::Waymo’s 5nm custom chip confirms that leading AV programs now treat compute architecture as proprietary IP, not a procurement decision.

⚡Humanoid Supply Chain::Schaeffler’s 2027 mass-production commitment for strain wave gearboxes is the clearest sign yet that the humanoid hardware supply chain is industrializing on schedule.

⚡LLMs in AV::Survey evidence confirms LLMs are being integrated as open-vocabulary reasoning layers in AV stacks — hybrid LLM-plus-deterministic architectures are becoming the new standard.

⚡Deployment Inflection::If Ouster’s demo-to-deployment call is accurate, the physical AI investment thesis is shifting from “when will this work?” to “who owns the supply chain when it does?” — and that question is still wide open.