🧠 Gemini Robotics 2 Gives Robots Full-Body Reasoning — DeepMind’s Most Capable Model Yet

• ●Google DeepMind’s Gemini Robotics 2 enables robots to reason through every individual movement across the full body, unlocking a significantly broader task range than its predecessor.
• ●The model integrates multimodal reasoning directly into whole-body motor control — a tighter coupling of perception, language, and action than prior VLA architectures achieved.
• ●Apptronik is among the hardware partners; watch for deployment announcements on humanoid platforms as DeepMind moves from research demos toward real-world validation.
• ●🔎 Read More →
• What matters: Full-body reasoning in a single model is the missing link between impressive robot demos and robots that can actually generalize across tasks.

🧪 Technology, Research & Innovation

🚗 Hierarchical RL Teaches Autonomous Cars to Drift — and Cut Lap Times

• ●Researchers propose a track-guided hierarchical RL framework that autonomously executes rally-style drifting — intentionally breaking traction on loose surfaces to minimize lap time.
• ●The dual-objective control problem (simultaneous drift stabilization and minimum-lap-time planning) is decomposed hierarchically, with a high-level planner feeding reference lines to a low-level drift controller.
• ●Solving this at the policy level — rather than via hand-tuned MPC — opens a path to AV systems that handle extreme, traction-limited dynamics without explicit physics models.
• ●🔎 Read More →
• What matters: Teaching an AV to drift on demand is a stress test for any control stack — success here signals robustness at the edge of the vehicle’s physical envelope.

🧠 Humanoid Learns to Walk, Reach, and Grasp by Composing Developmental Policies

• ●A new arXiv paper presents a developmental-learning approach where a humanoid first acquires a whole-body reaching-and-grasping policy, then composes it with a standing-up-and-walking policy into a unified mobile manipulation behavior.
• ●The interaction representation framework allows each sub-policy to remain independently trainable while the composition layer handles coordination — avoiding the combinatorial blowup of end-to-end joint training.
• ●If the composition generalizes across morphologies, this could become a standard recipe for bootstrapping humanoid skill libraries without retraining from scratch for every new task combination.
• ●🔎 Read More →
• What matters: Modular policy composition — not monolithic end-to-end training — may be the practical path to humanoids that can combine locomotion and manipulation reliably.

🚀 Product, Hardware & Model Launches

🧠 Fujitsu and NVIDIA Announce Joint Physical AI Push

• ●Fujitsu and NVIDIA are jointly advancing physical AI, combining Fujitsu’s enterprise infrastructure and domain expertise with NVIDIA’s accelerated computing and AI platform.
• ●The partnership positions NVIDIA’s physical AI stack — including Isaac and Cosmos — deeper into industrial and enterprise robotics deployments via Fujitsu’s established customer base.
• ●For the market, this is a signal that physical AI is moving from pilot projects into enterprise procurement cycles, with hyperscaler-tier compute backing the rollout.
• ●🔎 Read More →
• What matters: When Fujitsu and NVIDIA align on physical AI, it marks the transition from research curiosity to enterprise infrastructure category.

🤖 8 State-of-the-Art Humanoid Policies Fail Single-Leg Balance on All 90 Test Motions — New Benchmark Exposes the Gap

• ●A new benchmark tests eight released state-of-the-art general humanoid policies on single-leg balance across 90 motions — every policy scores zero clean stances, relying on stepping or hopping to recover rather than prevent imbalance.
• ●The paper introduces Dynamic Center-of-Mass (Dynamic-CoM) control as the first deployable method to pass the benchmark, and frames the evaluation as method-agnostic so any future policy can be compared on equal footing.
• ●This benchmark will pressure humanoid teams to explicitly target static balance primitives — a capability gap that becomes critical the moment robots operate on uneven terrain or interact with humans at close range.
• ●🔎 Read More →
• What matters: Zero out of 90 is a damning result — today’s best humanoid policies are optimized for agility, not stability, and that trade-off has real deployment consequences.

💰 Business, Startups & Investment

🧠 Ex-DeepMind Applied Robotics Leaders Launch Reimagine Robotics — No Programmers Required

• ●Reimagine Robotics has emerged from stealth, founded by former Google DeepMind Applied Robotics leaders, with a stated goal of building robots that learn on the job without requiring specialized programmers.
• ●The company targets manufacturing, assembly, and machine-tending applications — sectors where programming overhead has historically blocked cobot adoption at smaller facilities.
• ●With DeepMind pedigree and a zero-programmer pitch, Reimagine is positioning directly against the incumbent cobot model; watch for funding details and first customer announcements as the stealth lift continues.
• ●🔎 Read More →
• What matters: Eliminating the programmer requirement is the unlock that makes industrial robotics accessible to the long tail of manufacturers who can’t afford integration teams.

🚗 China’s Momenta Secures Nationwide Level 4 Testing Permit in Germany — First Chinese AV Firm to Do So

• ●Momenta has won a nationwide Level 4 autonomous vehicle testing permit in Germany, making it the first Chinese AV company to gain country-wide testing authorization in a major European market.
• ●Germany’s permit framework requires demonstrated safety cases and regulatory engagement — clearing this bar gives Momenta a credibility signal that matters for European OEM partnerships and future commercial licensing.
• ●As Waymo dominates U.S. headlines, Momenta’s European foothold widens the robotaxi race geographically and raises the competitive pressure on European AV incumbents like Mobileye and Wayve.
• ●🔎 Read More →
• What matters: A Chinese AV firm holding a nationwide German L4 permit redraws the competitive map — the robotaxi race is no longer a U.S.-China bilateral contest.

⚡Full-Body AI::Gemini Robotics 2’s whole-body reasoning marks a qualitative shift — the question is no longer whether foundation models can control robots, but how fast they generalize to unstructured environments.

⚡Balance Gap::Eight leading humanoid policies failing single-leg balance on 90/90 test cases is a hard data point that the field’s agility-first training paradigm has a structural blind spot.

⚡Policy Composition::Developmental skill composition — train sub-policies independently, compose at runtime — is emerging as a credible alternative to monolithic end-to-end humanoid training.

⚡Enterprise Physical AI::The Fujitsu-NVIDIA partnership signals physical AI is entering enterprise procurement cycles, not just research labs — infrastructure deals will precede mass robot deployments.

⚡Geopolitical AV Race::Momenta’s German L4 permit forces a question the industry has avoided: if Chinese AV firms clear European regulatory bars, does the West’s assumed safety-and-trust advantage still hold?