The day in conversation for 2026-10-29

discussion

Transcript

The discussion: Rosa: That Kinematic-Tactile Aligned Dexterous Retargeting work seems really promising regarding grounding actions in physical contact; I'm interested in how they managed the alignment of those contact wrenches specifically when dealing with non-rigid objects, because that’s where my simulations always struggle to capture the necessary compliance. Dev: I agree that physical alignment is crucial, but I wonder if we're overemphasizing just the grasping mechanics right now; looking at ScheduleGen, it feels like we might be solving *how* to grasp perfectly, but not yet *how efficiently* a multi-arm system can sequence those grasps based on task demands. What does that schedule generation actually imply about reducing the required sample size for imitation learning? Taro: I think the implication is that if ScheduleGen can generate high-fidelity demonstrations for complex sequences, we might bypass the brute-force exploration of skill space that usually plagues traditional reinforcement learning setups. However, my concern is generalization; even if we perfect one geometry in a controlled lab setting, how does a robot reliably translate that fine-grained tactile feedback strategy when the object geometry shifts drastically in an uncontrolled environment? Rosa: That shift in environment is exactly what worries me about those tactile skills. When we look at SimTO-Robust, it seems they relied heavily on pre-training against synthetic variations, but that doesn't account for the noise inherent in real-

More episodes

← Home