The day in conversation for 2026-11-08

discussion

Transcript

The discussion: Rosa: That 90-to-100% success rate on the Social Humanoid paper using the VQ-VAE codebook is impressive because it suggests a surprisingly low divergence between intended human motion and robot execution when using that shared latent space. What I find interesting, though, is how they manage the retargeting process when those motions are generated in simulation versus being executed on physical hardware with real-world latency constraints. Dev: The success rate depends heavily on the scenario complexity; if we test it against a very sparse or highly dynamic environment, that 90% might drop because the shared codebook might lack sufficient fine-grained motion primitives for those specific edge cases. Have you looked at how they handled the cross-modal transfer when moving from simulation to real robot control? Taro: I haven't deep-dived into the implementation details of the retargeting yet, but from a planning perspective, I’m more concerned about FlowDrive. Reaching state-of-the-art on nuPlan benchmarks by using data balancing to mitigate long-tail biases is a significant leap because those biases usually manifest as catastrophic failures in edge cases that standard reinforcement learning just ignores. Rosa: Exactly, the long-tail problem is huge. If we can reliably temper those biases in the planning stage like FlowDrive proposes, it means the resulting trajectories are more robust across different sensor inputs, which should theoretically make integrating complex perception systems easier down the line. Does that improved planning capability necessarily mean less reliance on highly specific motion generation techniques

More episodes

← Home