Why Robotics Still Isn't Solved - But Could Be Soon | YC Paper Club

In this talk

1. 0:25

   Why robotics remains difficult

   The host argues that repeated predictions of near-term breakthroughs have not made general-purpose robots practical. The obstacles include models that fail to capture real-world physics, weak representations for deformable objects and actions, limited tactile sensing, and hardware behavior that changes with wear. Teleoperation is also difficult to scale, and collected demonstrations can become stale as robots drift.

2. 8:00

   Memory for longer robot tasks

   Marcel presents Multi-Scale Embodied Memory, which combines dense visual history for low-level dexterity with a compact text scratchpad for high-level task progress. In demonstrations, memory helps robots avoid endlessly washing dishes or burning grilled cheese, and supports tasks lasting tens of minutes. It also helps policies notice repeated mistakes and change their approach; the text memory was supervised with annotated data.

3. 20:22

   Learning which reasoning helps action

   Milan describes R&B Encore, a self-supervised method for proposing and scoring reasoning traces for embodied models. Its criteria include brevity, useful content, and whether the trace predicts actions. Across manipulation, locomotion, navigation, and driving, the team found that useful reasoning depends on the embodiment: movement and gripper position helped manipulation, while some visual details and counterfactuals could distract or add redundancy.

4. 33:43

   Dexterity from simulation and practical deployment

   Tyler Lum explains Sim-to-Real, which trains a single dexterous hand-and-arm policy with parallel reinforcement learning in simulation, then uses human videos only to provide goal poses at deployment. The frozen policy handled 12 unseen tools and tasks, though heavy tools, small objects, and pose-tracking errors limited performance; about 60% of failures were attributed to tracking. Nico argues that robotics companies can start with one paid customer problem, teleoperation, and off-the-shelf hardware, then improve through rapid data collection, evaluation, and operational learning.

5. 1:08:30

   Making world action models faster

   Bill and Guanmin of General Instinct describe world action models that learn from predicted future video and robot actions. Because diffusion-based video prediction can be computationally costly, they discuss keeping useful representations in hidden states, shrinking models, and distilling flow-matching inference to one or two steps. They report a runtime of 500 milliseconds per 16-action chunk on Jetson Orin.