The day in conversation for 2026-11-03

discussion

Transcript

The discussion: Rosa: The improvement in feasibility for quadrotor trajectory generation from that convex reformulation sounds substantial; how does that exact formulation handle the discrete-time nature of the stochastic systems compared to what we saw in our earlier work on pure continuous control? Dev: I’m less focused on the exact math and more on how this affects online adaptation. If we can formulate constraints exactly, does that translate into a faster convergence rate when the system dynamics are changing unexpectedly mid-flight? Taro: That ties directly into the RL side, actually. When we look at pi-StepNFT, they are coupling flow-native exploration with finer step-wise supervision. It seems like this might be how we bridge the gap between those deterministic trajectory improvements and handling the inherent noise in real robotic execution. Rosa: I wonder if that coupling is robust enough for truly high-dimensional contact tasks, which is what we keep coming back to as the bottleneck. Does their supervision rely on a dense set of demonstrations, or can it generalize from sparse initial input? Dev: The paper mentions success rates on LIBERO and ManiSkill, implying they are working with existing benchmarks where data availability might be constrained. If the flow-native framework can maintain stability when transitioning between those supervised steps, that’s a huge step away from purely model-based planning approaches. Taro: That brings up the LfD work too. Decompose and Reorganize seems to tackle the high-dimensional problem by breaking policies into atomic skills derived from demonstrations. Is the challenge with integrating those decomposed skills into a single, unified motion plan,

More episodes

← Home