A B-spline Policy UNet diffusion policy trained on Dimios45/yam-pick-duster-200: 200 teleop demos of an I2RT YAM arm picking a blue duster and placing it in a red box.

The policy predicts B-spline knots and control points instead of a fixed grid of actions, giving a continuous trajectory that can be resampled at any rate and temporally rescaled at deploy time.

Commands joints directly — no inverse kinematics at inference. This is the model to deploy. The end-effector counterpart trained on the same 200 takes is Dimios45/yam-pick-duster-200-bspline-ee, but its dataset card advises against deploying that Cartesian path unresolved.

Supersedes the 50-episode yam-pick-duster-bspline-joint.

action / obs joint_pos  (7,) = [joint1..6 (radians), gripper]
gripper: 0 = open, 1 = closed

Two observation steps; random crop to 76x76 in training, center crop at eval.

Actions stay 7-D: with action.shape [7], _convert_actions takes the raw_dim == target_action_dim branch and passes them through untouched — no rotation_6d expansion. infer_action_meta resolves this to single_yam_joint.

Network output is (16, 8): column 0 is the knot vector in units of 25 Hz frames relative to the current observation, columns 1-7 are control points. 16 = chunk_size 10 + 2 x degree 3.

Both embed the Hydra config (pickled with dill), so bspline_policy and diffusion_policy must be importable on load.

Loss: 0.167 (ep0) -> 0.007 (50) -> 0.005 (100) -> 0.003 (150) -> 0.002 (200-250), final in-epoch 0.00124.

The 251-epoch budget was chosen on gradient-step count (~300k steps, 1.13x the 50-episode run that converged cleanly), not copied from it — 601 epochs on this data would have been ~9.5 h for no benefit. checkpoint_every: 50 with 251 (not 250) epochs so the fully-annealed final epoch is actually written; saves land only on multiples of 50.

At matched epochs this beats the 50-episode model: 0.005 vs 0.007 at epoch 100.

Measured on an idle GPU, batch 1, two 84x84 cameras:

A chunk spans ~0.4 s of wall time at 25 Hz, and inference must finish inside it. On GPU all three settings clear comfortably. On a NUC-class CPU (expect 2-3x the CPU column), 16 steps is tight — use --num-inference-steps 8 on CPU. Set --predict-before-end to roughly 2-3x your chosen latency.

Requires the joint-space deployment patches (single_yam_joint decoder, yam_server joint mode bypassing IK, the top_image camera). See Dimios45/yam-duster-bspline-dp for the full list — they are not in upstream B-spline-policy/bspline-policy.

Trained uncropped (RAW 640x480 -> 84x84). The dataset card recommends --crop top_image=42,28,598,414; this model did not use it, and whatever crop you train with must be applied identically at deployment.

Gripper is 0 = open, 1 = closed. Verify on hardware before the first rollout — if the convention is reversed the policy opens to grasp and closes to release, which looks almost-working and is hard to diagnose:

Expect ~0.0 open, ~1.0 closed.

All 200 episodes were kept, including the 2 aborted takes that never close the gripper (198/200 contain a grasp) and the 14 with recovery re-grasps.