Commentary on X
@Haoyu_Xiong_ Had fun reading the paper! I did B-Splines for supervisory UAV flight. I agree flight envelope/actuator constraints are the hard limit. Do you think feeding back tracking error could make the policy more adaptive? Of course the policy had to be e2e embedded with the PIDs.
9:23 AM · Jul 15, 2026@Haoyu_Xiong_ This is awesome to see. Let's connect. We are exploring the same idea but at data gathering time. Would love to get your input:))
8:26 AM · Jul 15, 2026@Haoyu_Xiong_ man b-splines make so much sense to me, cool work!!
10:35 AM · Jul 15, 2026@Haoyu_Xiong_ Exciting to see! Would you be releasing code?
9:19 AM · Jul 15, 2026View all
How can we speed up robot policies? Excited to share B-spline Policy, where we predict future actions as a continuous parametric curve. This allows us to rapidly speed up manipulation policies by changing the rate time steps are sampled from the curve.
7:58 AM · Jul 15, 2026Throughput is actually the right metric for robot evals if you want to automate a task (speed * success rate)
8:32 AM · Jul 15, 2026