Physical AI, AD and Robotics

Newsletter | Technical Briefing

Curated insights for AD professionals, Roboticists, Physical-AI engineers, Founders & Tech leaders

🧠 SceneBot tracks humanoid contact with objects and terrain in real-time

• ●Researchers introduced SceneBot, a unified motion-tracking framework that handles contact-rich tasks like object manipulation and uneven terrain navigation—capabilities current humanoid RL policies lack because pure kinematic tracking can’t resolve physical ambiguities during interaction.
• ●The system bridges the gap between free-space locomotion (where existing policies excel) and contact-heavy scenarios by incorporating scene interaction directly into whole-body tracking, enabling humanoids to reason about forces and constraints.
• ●This approach could unlock deployment of humanoid policies in real-world manufacturing and logistics environments where contact with objects, tools, and irregular surfaces is unavoidable.
• 🔍 Read More →
• What matters: Humanoid policies can now track and learn from contact-rich interactions, not just free-space motion.

🧠 Direct action-head injection of grounded 3D points unlocks spatial and task generalization in VLA models

• ●Vision-Language-Action models trained on large-scale data remain brittle when object positions shift from training configurations or when familiar scenes require different language instructions—two critical failure modes limiting real-world deployment.
• ●Researchers developed a method that injects grounded 3D point coordinates directly into the action head, allowing the model to spatially ground instructions and generalize across both novel object placements and task variations without retraining.
• ●Early results suggest this architectural change could reduce the data requirements for teaching robots new tasks in existing environments, a key bottleneck for commercial manipulation systems.
• 🔍 Read More →
• What matters: VLA models can now handle spatial and task variations without expensive retraining cycles.

🤖 Support-constrained RL improves real-world policies using only simulation

• ●Robots trained on real-world data are often slow and brittle, but improving them with RL typically requires costly real-world training; unconstrained sim-based RL, meanwhile, produces policies that fail when deployed due to sim-to-real gaps.
• ●This work introduces support-constrained RL, which restricts policy updates to stay within the distribution of real-world data while still allowing improvement through simulation—essentially using sim as a safe sandbox for refinement rather than a replacement for real experience.
• ●The method enables iterative policy improvement without additional real-world data collection, potentially accelerating deployment cycles for manipulation and mobile manipulation systems.
• 🔍 Read More →
• What matters: Policies can now be refined in simulation without the distribution shift that typically breaks sim-to-real transfer.

🤖 SimFoundry automates zero-shot real-to-sim scene construction from video

• ●Training and evaluating robot policies in the real world is expensive and hard to scale; SimFoundry addresses this by automatically generating sim-ready digital twins from a single video, complete with object, scene, and task editing capabilities.
• ●The system supports automated generation of diverse digital copies of real environments, enabling policy learning and evaluation at scale without physical infrastructure or manual scene modeling.
• ●This tooling could compress the iteration cycle for manipulation research, allowing teams to test policies across hundreds of scene variations before committing to real-world trials.
• 🔍 Read More →
• What matters: Real-world scenes can now be digitized and varied automatically, removing a major bottleneck in policy development.

🤖 Generative video priors unlock expressive quadruped motion beyond animal gaits

• ●Quadruped robots remain confined to a handful of gaits despite advances in locomotion; this work bypasses the traditional pipeline of collecting animal motion data by using generative video priors to synthesize diverse, expressive behaviors directly.
• ●The approach challenges the assumption that robot motion must first pass through an animal body—instead, it generates motion that is physically plausible for the robot’s morphology while dramatically expanding the behavioral repertoire.
• ●If validated on hardware, this could enable quadrupeds to perform companion-like, context-appropriate behaviors in human environments, moving beyond utilitarian locomotion toward social robotics applications.
• 🔍 Read More →
• What matters: Quadrupeds may soon move beyond gaits, unlocking expressive behaviors without animal motion capture.

🤖 Proception settles Tesla trade secret suit and closes $11M Series A

• ●Proception, a startup building dexterous robot hands, settled a trade secret lawsuit with Tesla and announced an $11M Series A, signaling investor confidence despite the legal overhang.
• ●The company is taking a unique approach to one of robotics’ hardest problems—dexterous manipulation—by focusing on novel training data collection methods that could accelerate learning for contact-rich tasks.
• ●The funding and legal resolution position Proception to scale hardware production and data collection infrastructure, critical steps toward commercial deployment of general-purpose manipulation systems.
• 🔍 Read More →
• What matters: Dexterous manipulation startups are attracting capital despite technical and legal risks.

🤖 BMW deploys Figure 03 humanoid after Figure 02 supported production of 30,000 vehicles

• ●BMW’s Figure 02 humanoid supported production of over 30,000 BMW X3 vehicles across 11 months at the South Carolina plant, demonstrating sustained real-world utility in automotive manufacturing.
• ●The automaker is now deploying the upgraded Figure 03 model, suggesting the initial deployment met performance and reliability thresholds required for continued investment in humanoid automation.
• ●This marks one of the longest documented humanoid deployments in high-volume manufacturing, providing a critical data point for other OEMs evaluating humanoid ROI against traditional automation.
• 🔍 Read More →
• What matters: Humanoids are moving from pilots to sustained production roles in automotive manufacturing.

⚡Contact-rich manipulation::Humanoid and manipulation policies are finally addressing the contact problem—tracking and learning from physical interaction, not just free-space motion.

⚡Sim-to-real iteration::New methods for policy refinement in simulation and automated scene digitization are compressing the cost and time required to improve real-world robot performance.

⚡VLA generalization::Architectural innovations like grounded 3D point injection are making vision-language-action models more robust to spatial and task variations without retraining.

⚡Humanoid manufacturing::BMW’s 30,000-vehicle production run with Figure 02 provides the first large-scale validation of humanoid ROI in automotive assembly.

⚡Beyond gaits::If generative video priors can unlock expressive quadruped motion at scale, the line between utilitarian and social robotics may blur faster than the industry expects.

Worth forwarding to a colleague? Pass it along.

© 2026 Physical AI, AD and Robotics · DriveTech AI. All rights reserved. Privacy Policy