🚗 Tesla Robotaxi Crosses 1 Million Unsupervised Miles

• ●Tesla has confirmed its Robotaxi has completed 1 million miles of fully unsupervised operation — a concrete fleet milestone that puts it in direct comparison with Waymo’s driverless program.
• ●Unsupervised miles are the hardest metric to fake: no safety driver, no remote intervention, meaning Tesla’s FSD stack is now handling edge cases at scale without a human fallback.
• ●Watch for Tesla to use this figure as regulatory leverage in new markets and as a commercial launch signal — the next milestone to track is revenue per mile from paid rides.
• ●🔎 Read More →
• What matters: One million unsupervised miles reframes Tesla from an ADAS company into a genuine robotaxi operator — the regulatory and commercial clock just accelerated.

🧪 Technology, Research & Innovation

🧠 XR-2 VLA Trained on 1,500 Hours of Bimanual Household Demos

• ●Researchers released 1,500 hours of diverse bimanual manipulation demonstrations covering everyday household tasks — currently the largest dataset of its kind for two-armed robot learning.
• ●The corpus trains XR-2, a vision-language-action model, then applies on-policy corrections to close the sim-to-real gap that typically degrades generalist policies at deployment.
• ●The open dataset release is the real story: it sets a new data-scale baseline that other VLA teams will need to match or surpass to claim competitive bimanual performance.
• ●🔎 Read More →
• What matters: 1,500 hours of labeled bimanual data plus on-policy correction is the clearest path yet to household robots that don’t fail on the second attempt.

🚗 New World Model Tackles Long-Horizon Consistency for End-to-End Autonomous Driving

• ●A new arXiv paper identifies three failure modes crippling current world-model-based RL for autonomous driving: temporal inconsistency over long rollouts, poor ego-environment interaction modeling, and inability to handle multi-style driving behavior.
• ●The proposed framework enforces interaction-aware rollouts — meaning the simulated world updates in response to the ego vehicle’s actions, not just time — which is critical for accurate credit assignment during RL training.
• ●Multi-style support (aggressive, conservative, highway, urban) in a single model is the differentiator to watch; it suggests a path toward one policy that passes diverse regulatory test scenarios.
• ●🔎 Read More →
• What matters: Fixing temporal inconsistency in imagined rollouts is the unglamorous bottleneck that separates world-model RL from real-world AV deployment — this paper attacks it directly.

🚀 Product, Hardware & Model Launches

🧠 BRIDGE: Open-Source Humanoid Platform Co-Designs Hardware and Whole-Body Control Together

• ●BRIDGE is a fully open-source humanoid robot platform built around morphology-control co-design — hardware geometry and whole-body controller are optimized jointly rather than sequentially, a departure from standard practice.
• ●The co-design approach directly targets the fluidity gap: conventional humanoids are stiff and jerky because actuator placement is finalized before the controller is tuned, forcing the software to compensate for hardware it didn’t influence.
• ●Open-sourcing the full stack — CAD, control code, and training pipeline — means academic labs can now iterate on humanoid morphology without a multi-million-dollar hardware budget.
• ●🔎 Read More →
• What matters: BRIDGE democratizes humanoid research the way ROS democratized manipulation — co-design as a first principle could redefine what open-source embodied AI looks like.

🤖 NASA’s CADRE Mission Will Deploy Three Autonomous Rovers on the Moon with No Pre-Assigned Tasks

• ●NASA’s CADRE (Cooperative Autonomous Distributed Robotic Exploration) mission will land three small rovers on the Earth-facing lunar surface with a single collective instruction: self-organize and explore a designated patch of ground.
• ●The mission tests multi-agent coordination without ground-in-the-loop commands — communication latency to the Moon makes real-time human control impractical, so the rovers must negotiate task allocation and path planning autonomously.
• ●CADRE is effectively a live stress-test of decentralized swarm autonomy in a GPS-denied, comms-delayed environment — results will directly inform rover architectures for Mars and beyond.
• ●🔎 Read More →
• What matters: CADRE is the highest-stakes real-world test of autonomous multi-robot coordination ever attempted — the Moon doesn’t offer a second chance to debug.

💰 Business, Startups & Investment

🧠 NVIDIA Acquires Hugging Face for $12.93 Billion

• ●NVIDIA has agreed to acquire Hugging Face for $12,930,300,000 — making it one of the largest AI infrastructure acquisitions on record and giving NVIDIA direct ownership of the dominant open-source model hub.
• ●Hugging Face hosts hundreds of thousands of models, datasets, and Spaces used daily by robotics and physical AI teams; NVIDIA now controls that distribution layer alongside its GPU compute stack.
• ●The critical question for the robotics ecosystem: whether NVIDIA keeps Hugging Face’s open-access model or shifts it toward a compute-tied, platform-locked offering that favors NVIDIA hardware.
• ●🔎 Read More →
• What matters: NVIDIA just bought the app store for AI models — whoever controls model distribution at this scale shapes which physical AI architectures get built next.

🚗 PlusAI Goes Public via SPAC at $800 Million Valuation

• ●PlusAI, whose autonomous trucking software is already generating commercial revenue, has agreed to a SPAC transaction valuing the company at approximately $800 million.
• ●Unlike most AV SPAC deals of the early 2020s, PlusAI enters the public markets with a revenue-generating product — its software runs on commercial freight trucks today, not in a test fleet.
• ●The $800M valuation will be the benchmark: if PlusAI trades above it post-merger, it signals renewed public-market appetite for autonomous trucking; a discount would confirm lingering AV skepticism.
• ●🔎 Read More →
• What matters: PlusAI’s revenue-first SPAC is the autonomous trucking sector’s best shot at proving public markets can price AV companies on fundamentals, not promises.

⚡Robotaxi Reality::Tesla’s 1M unsupervised miles shifts the AV conversation from capability demos to operational scale — regulators and competitors must now respond to a real number.

⚡Data as Infrastructure::The XR-2 dataset release proves that 1,500 hours of quality bimanual demos is now the minimum ante for credible generalist manipulation research.

⚡Open Hardware::BRIDGE’s co-design methodology challenges the assumption that humanoid development requires proprietary, closed hardware — open-source morphology is now a viable research path.

⚡Distribution Control::NVIDIA’s $12.93B Hugging Face acquisition means the company now owns compute, models, and the platform where physical AI teams share and discover both.

⚡The Autonomy Stack Consolidates::From CADRE’s swarm rovers to PlusAI’s public debut, the week’s news points to one question worth debating: as autonomy matures from research to infrastructure, does openness survive consolidation — or does it become the first casualty?