🚗 NHTSA Opens Probe Into Tesla’s Self-Certification of Cybercab Robotaxis

• ●The US National Highway Traffic Safety Administration launched a formal investigation into whether Tesla properly self-certified the Cybercab under federal motor vehicle safety standards — a process that bypasses traditional regulatory review.
• ●Self-certification is legal under US law, but NHTSA’s probe signals scrutiny of whether a vehicle designed without a steering wheel or pedals meets existing FMVSS rules written for human-operated cars.
• ●The outcome will set a precedent for how purpose-built robotaxis — from any manufacturer — navigate a regulatory framework that has not yet been formally updated for fully driverless vehicles.
• ●🔎 Read More → Reuters
• What matters: A federal probe into Cybercab self-certification is the first real regulatory stress-test of the US framework for purpose-built, driverless vehicles.

🧪 Technology, Research & Innovation

🧠 NVIDIA NemoClaw Gives AI Agents Persistent Memory Across Tasks

• ●NVIDIA’s NemoClaw framework enables AI agents to store, retrieve, and act on episodic and semantic memory, moving beyond single-session context windows that limit most current LLM-based agents.
• ●The architecture separates memory into distinct stores — short-term working memory and long-term retrieval — allowing agents to personalize behavior and maintain task continuity across interactions without retraining.
• ●For physical AI applications, persistent memory is a prerequisite for robots and autonomous systems that must adapt to user preferences and evolving environments over days or weeks, not just single sessions.
• ●🔎 Read More → NVIDIA
• What matters: NemoClaw’s memory architecture is a concrete step toward AI agents that accumulate operational knowledge — a capability physical AI systems need before they can be genuinely useful in unstructured environments.

🚗 Physics-Native AI Could Sharpen AV Sensors and Biomedical Imaging Simultaneously

• ●Researchers have developed an AI model trained directly on physical signal data — radar, ultrasound, and lidar waveforms — rather than on image representations, improving fidelity at the raw-signal level.
• ●Processing signals before image conversion preserves phase and amplitude information that conventional vision-based AI discards, which matters for AV sensors operating in rain, fog, or low-light conditions where image quality degrades first.
• ●The dual applicability to biomedical imaging and AV sensing suggests a shared research path — advances in one domain could directly accelerate the other, compressing development timelines for both.
• ●🔎 Read More → Technology Org
• What matters: AI built for physical signals — not images — could give AV sensors a meaningful edge in exactly the adverse conditions where current perception systems are most likely to fail.

🚀 Product, Hardware & Model Launches

🧠 NVIDIA Cosmos 3 Lands on AWS SageMaker HyperPod as a Physical AI Model Factory

• ●AWS has integrated NVIDIA Cosmos 3 into SageMaker HyperPod, enabling teams to train, fine-tune, and deploy world-model-based physical AI pipelines directly on managed cloud infrastructure without standing up custom clusters.
• ●Cosmos 3’s world-model architecture generates physically plausible synthetic training data for robotics and AV systems — running it on HyperPod means teams can scale data generation and model training in the same managed environment.
• ●The AWS partnership lowers the infrastructure barrier for mid-sized robotics and AV teams that need Cosmos-scale compute but lack the resources to operate dedicated GPU clusters on-premise.
• ●🔎 Read More → Amazon Web Services (AWS)
• What matters: Cosmos 3 on HyperPod turns AWS into a turnkey physical AI model factory — making world-model-scale training accessible without dedicated on-premise GPU infrastructure.

🤖 Lattice FPGAs Positioned as Deterministic Security Guardrails for Physical AI Systems

• ●Lattice Semiconductor’s VP of security Eric Sivertson argues that FPGAs — not software layers — are the right enforcement point for physical AI security, because their deterministic execution cannot be overridden by a compromised host processor.
• ●In robotic systems, FPGAs can enforce hard safety envelopes at the hardware level — intercepting commands from AI inference engines before they reach actuators — providing a trust boundary that software-only approaches cannot guarantee.
• ●As humanoids and AMRs move into manufacturing and public spaces, hardware-enforced safety boundaries are becoming a procurement and liability requirement, not just a design preference — positioning FPGA vendors as critical infrastructure suppliers.
• ●🔎 Read More → The Robot Report
• What matters: Hardware-enforced safety boundaries via FPGAs may become a non-negotiable layer in any physical AI system deployed where a software failure has physical consequences.

💰 Business, Startups & Investment

🚗 Horizon Robotics CEO Targets NVIDIA’s High-End AV Chip Crown in China by 2027

• ●Horizon Robotics founder Yu Kai publicly stated the company aims to surpass NVIDIA in China’s high-end autonomous driving chip market within the next year — a direct challenge to NVIDIA’s Orin and Thor SoC dominance in the segment.
• ●Horizon’s Journey series chips are already designed into vehicles from major Chinese OEMs; the high-end push targets the compute-intensive perception and planning workloads where NVIDIA currently commands premium pricing and design wins.
• ●US export controls on advanced NVIDIA chips to China have created a structural opening for domestic alternatives — Horizon is the best-positioned local player to fill it, but execution at scale remains the test.
• ●🔎 Read More → finance.biggo.com
• What matters: US export controls have turned China’s AV chip market into a domestic race — and Horizon Robotics is betting it can out-execute NVIDIA on its home turf within 12 months.

🤖 Lyte AI Raises $165M to Scale Robot Perception — Its Second Round This Year

• ●Lyte AI closed a $165M funding round — its second raise in 2026 — to scale production of its AI-driven perception capabilities designed to help robots better sense and interpret their physical surroundings.
• ●Two large rounds in a single year signals investor conviction that perception remains the binding constraint on robot deployment, and that Lyte’s approach — combining AI inference with sensor fusion — is differentiated enough to warrant aggressive scaling.
• ●Watch whether Lyte moves toward vertical integration with hardware, or positions itself as a perception software and module supplier to humanoid and AMR OEMs — the go-to-market choice will define its competitive moat.
• ●🔎 Read More → The Robot Report
• What matters: Two $100M+ rounds in one year for a robot perception startup confirms that sensing — not actuation or compute — is where investors currently see the largest unsolved gap in physical AI.

⚡Regulatory Reckoning::NHTSA’s Cybercab probe is the first real test of whether US self-certification rules can hold for purpose-built driverless vehicles — the outcome shapes every robotaxi program’s compliance roadmap.

⚡Memory as Infrastructure::NVIDIA NemoClaw treats persistent memory as a first-class system component — a design choice that will become standard in any physical AI agent expected to operate across sessions.

⚡Cloud-Scale Physical AI::Cosmos 3 on AWS HyperPod democratizes world-model training, removing the on-premise GPU cluster requirement that previously limited physical AI development to well-capitalized labs.

⚡Hardware Security Layer::FPGAs as deterministic safety guardrails represent a maturing view of physical AI architecture — one where trust boundaries are enforced in silicon, not software.

⚡The Perception Bet::With Lyte’s $165M second round and Horizon’s NVIDIA challenge, 2026 is shaping up as the year the industry decides whether perception is a commodity layer or the defining competitive moat — and the answer will restructure the entire physical AI supply chain.