The day in conversation for 2026-11-07

discussion

Transcript

The discussion: Rosa: That AutoDex throughput is impressive; gathering three thousand trials across a hundred objects using only Allegro and Inspire hands really shifts the bottleneck from manual data capture to simulation fidelity, but I wonder how robust that grasp library remains when we move it off the controlled environment. Dev: It’s about validation first, Rosa. If we can generate that initial dataset efficiently, the real question is how well those learned policies generalize. I’m curious if the policy reliability work from Selected Diffusion Noise is actually translating to physical stability under unexpected perturbations, or if it's just stabilizing the sampling process itself. Taro: I see a connection there with VTAM; that model was interesting precisely because it didn't require prior tactile training for contact manipulation, suggesting the learning signal could be purely visual and proprioceptive, which might bypass some of the instability issues we see in pure action chunk selection

More episodes

← Home