Dimios45/yam-pick-duster-200-bspline-ee

YAM Pick-Duster 200 — B-spline Diffusion Policy (end-effector)

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

This model emits absolute Cartesian poses and requires IK on the robot. The joint-space counterpart — same 200 takes, no IK, and the one recommended for deployment — is Dimios45/yam-pick-duster-200-bspline-joint. The pair exists to compare action spaces on identical demonstrations.

Read this before deploying. The dataset card states these poses are MuJoCo FK and disagree with bspline-policy's pyroki/URDF by up to ~9 mm, and advises: "do not deploy a policy trained on this data through its Cartesian path without resolving the discrepancy — drive the arm in joint space instead." Their suggested mitigations are EEFollower with low mu, target_frame="tcp", and reach clamping.

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

Action and observation space

raw actions (7,) = [pos(3), rotvec(3), gripper(1)]
   -> dataset expands to (10,) = [pos(3), rot6d(6), gripper(1)]
gripper: 0 = open, 1 = closed

key	shape	notes
wrist_image	(3, 84, 84)	RGB, RAW 640x480 resized — not cropped
top_image	(3, 84, 84)	RGB, overhead
arm_pos	(3,)	TCP position, metres
arm_quat	(4,)	xyzw, w >= 0 hemisphere
gripper_pos	(1,)	0 = open, 1 = closed

rotation_rep: rotation_6d with abs_action: True. infer_action_meta resolves this to single_yam_rot6d. Network output is (16, 11): column 0 is the knot vector in units of 10 Hz frames, columns 1-10 are control points.

TCP frame is the link_6 flange with a fixed 90 degree z-rotation, ~13.5 cm from the fingertip.

Files

file	size	use
deploy_ema.ckpt	426 MB	Inference. EMA weights only.
epoch0400_full.ckpt	1.5 GB	model + ema_model + optimizer, for resuming/fine-tuning.

Training

data	200 episodes, 30,462 frames @ 10 Hz -> 30,262 B-spline chunks
hardware	1x RTX 4090, ~2.6 h, 23 s/epoch, 473 batches/epoch
epochs / batch	401 / 64
optimizer	AdamW, lr 1e-4, cosine, 500 warmup, EMA
scheduler	DDIM, 100 train timesteps, 16 inference steps, epsilon prediction
B-spline	degree 3, chunk_size 10, max_error 0.002, absolute knots

Loss: 0.408 (ep0) -> 0.015 (50) -> 0.011 (100) -> 0.008 (150) -> 0.006 (200) -> 0.005 (250) -> 0.003 (300-350) -> 0.002 (400), final in-epoch 0.00157.

The 401-epoch budget was chosen on gradient-step count (~190k steps), not copied from the 50-episode run. checkpoint_every: 50 with 401 (not 400) epochs so the fully-annealed final epoch is actually written.

These numbers are not comparable to the joint-space model's: different action space, units, and normalizer. Only the shape of each curve is meaningful on its own.

Inference latency

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

DDIM steps	RTX 4090	CPU (i9-13900K, 8 threads)
4	13.4 ms	57.6 ms
8	24.2 ms	91.0 ms
16	45.8 ms	168.1 ms

A chunk spans ~1.1 s of wall time at 10 Hz, so all settings clear comfortably even on a NUC-class CPU (expect 2-3x the CPU column). 16 steps is a safe default here.

Rollout

cd ~/bspline-policy
export PYTHONPATH=$PWD/bspline_policy:$PWD/diffusion_policy:$PWD/real_env/yam_teleop

hf download Dimios45/yam-pick-duster-200-bspline-ee deploy_ema.ckpt --local-dir ./ckpt

# terminal 1 - arm server
sudo ip link set can_follower_r up type can bitrate 1000000
python real_env/yam_teleop/yam_server.py --channel can_follower_r

# terminal 2 - rollout
python real_env/yam_teleop/rollout_local_policy.py \
  --env yam-vrkit --policy bspline \
  --ckpt-path ./ckpt/deploy_ema.ckpt \
  --diffusion-policy-dir $PWD/diffusion_policy \
  --control-freq 100 \
  --data-freq 10 \
  --origin-time-scale 10 \
  --num-inference-steps 16 \
  --predict-before-end 0.3 \
  --speed-up-times 1.0 \
  --save --output-dir data/rollouts_ee_200

--env yam-vrkit is not in upstream bspline-policy. rollout_local_policy.py accepts only x5, tidybot2, and yam, so this fails at argparse until a matching env class is added. Substituting --env yam routes through RealEnv + yam_server's pyroki velocity IK — which is exactly the Cartesian path the dataset card warns about above.

Flags that are not optional

flag	value	why
--origin-time-scale	10	Knots are in data-frame units; must equal the training rate. The joint-space counterpart is 25 Hz — never copy this flag between the two models. Using 25 here runs the arm at 2.5x speed.
--data-freq	10	Must match the above.
--control-freq	100	Matches YAM_CONTROL_HZ.
--speed-up-times	start at 1.0	Velocity scales linearly, acceleration quadratically.
--predict-before-end	0.3	Must exceed inference latency.

A missing top_image key does not raise: policy_local_bspline.py:625-629 substitutes a black frame, so an unwired or dead camera yields a half-blind policy rather than an error.

Cameras, gripper, data notes

Trained uncropped (RAW 640x480 -> 84x84); reproduce exactly at deployment. Gripper is 0 = open, 1 = closed — verify on hardware first (see the joint-space card for the one-line check). All 200 episodes kept, including the 2 aborted takes and 14 with re-grasps.

Reproducing

cd real_env/yam_teleop && python convert_to_robomimic_hdf5.py \
  --input-dir ~/data/yam-200-ee --output-path ../../data/yam_ee_200.hdf5

cd bspline_policy && python train.py \
  --config-name=yam_pick_ee_200_bspline \
  hydra.run.dir=../outputs/yam_pick_ee_200 \
  training.resume=false logging.mode=offline \
  checkpoint.topk.k=999 dataloader.persistent_workers=True

Give this task config a cache_suffix distinct from every other run, or it will silently load a stale zarr cache.

Citation

@article{han2026b,
  title={B-spline Policy: Accelerating Manipulation Policies via B-spline Action Representations},
  author={Han, Xiaoshen and Xiong, Haoyu and Chen, Haonan and Liu, Chaoqi and
          Torralba, Antonio and Zhu, Yuke and Du, Yilun},
  journal={arXiv preprint arXiv:2607.09648},
  year={2026}
}