Four technical walls block scaling of robotics according to FrancoisFrancois outlines four key technical 'walls' preventing the scaling of robotics: physical real-world modeling (the sim-to-real gap), unsolved action space representation, the sensory-motor issue (lacking human-like tactile sensing), and embodiment drift (hardware degradation affecting learned policies).Francois ChaubardChallenges in robotics developmentSource at 2:06Save viewpoint
MAM system combines short-term visual and long-term textual memory for robots.Marcel presents MAM (multi-scale embodied memory), a system that decomposes memory into short-term visual memory for low-level dexterous control and long-term textual memory for high-level policy planning, enabling robots to track task progress and adapt to mistakes during long-horizon tasks.Marcel Torne VillasevilEmbodied memory for robot policiesSource at 7:59Save viewpoint
Simulation-trained policy achieves zero-shot real-world tool manipulationTyler describes SimToolReal, a method that trains a single, goal-conditioned policy in simulation using reinforcement learning to perform dexterous tool manipulation (grasp, reorient, use) zero-shot on real-world tools without task-specific retraining, demonstrating strong recovery behaviors.Tyler LumSim-to-real reinforcement learning for dexterous manipulationSource at 30:14Save viewpoint
R&B Encore uses self-supervised chain-of-thought reasoning for VLAs.Milan presents R&B Encore, a self-supervised pre-training cycle for VLAs that generates and validates embodiment-specific chain-of-thought reasoning, arguing that selective, action-predictive reasoning (like move and gripper position) is more critical than exhaustive or perceptual reasoning for diverse embodiments.Milan GanaiEmbodied reasoning in VLAsSource at 20:16Save viewpoint
Robotics application companies should own problems and build operational moats.Nico advocates for 'robotics application companies' that take full ownership of a business problem, start with teleoperation to prove viability, iterate rapidly with data tools, and build moats through operations and deployment, seeing this model as the new SaaS for the physical world.Nikolaus WestRobotics application companies and business modelsSource at 64:10Save viewpoint
World action models are computationally expensive, a barrier to deploymentBill and Guanming from General Instinct introduce infrastructure for optimizing world action models (WAMs), which predict future frames and kinematics simultaneously, highlighting their computational expense (requiring expensive GPUs) as a key scalability barrier for real-world deployment.Bill JiaoWorld action models and infrastructureSource at 83:52Save viewpoint
Distillation and separate transformers speed up world action model inferenceGuanming explains optimizations for WAMs, including distillation and separating the model into video and action transformers to avoid decoding future frames, significantly speeding up inference by maintaining world representation in hidden states and reducing flow matching steps.Guanming WangWorld action models and infrastructureSource at 85:31Save viewpoint
