Robotics Scaling Walls: Four Hard Problems Teams Still Face

The opening frames robotics' perennial 'next year' fallacy: despite diffusion policies and world models, commercial deployment remains elusive. Four scaling walls persist: physical world modeling, deformable objects, action representation, and the overlooked sensory-motor gap. The latter—humans' tactile ability vs robots' crude sensors—is a fundamental limitation.

Marcel's multiscale embodied memory (MAM) tackles long-horizon tasks by adding structured memory to policies, enabling in-context adaptation. The tradeoff: memory improves reliability but requires careful compression and annotation to avoid latency. Milan's self-supervised reasoning shows selective reasoning outperforms exhaustive reasoning, and reasoning must be action-predictive to be useful.

Tyler's sim-to-real RL achieves zero-shot generalization to real tools by training in simulation with domain randomization. The tradeoff: simulation cannot handle all real-world physics (deformable, water), but it avoids teleoperation bottlenecks. Nico's pitch: start robotics application companies by solving narrow problems with teleop first, then fine-tune.