Dimios45/yam-duster-dp

YAM Duster-in-Box — Diffusion Policy (baseline)

A UNet Diffusion Policy trained on Dimios45/yam_duster_in_box for a single-arm I2RT YAM robot: "pick up the duster and put the duster in the box."

This is the baseline for the B-spline Policy comparison. It was trained on byte-identical data with the same backbone, optimizer, schedule, and epoch count as Dimios45/yam-duster-bspline-dp — the only difference is the action representation. Use it to measure what the B-spline parameterization actually buys in success rate and completion time.

Action space

Joint-space targets, not end-effector poses:

[right_joint_1..6 (radians), right_gripper (0 = open, 1 = closed)]

No IK is involved at deploy. The policy predicts a plain (16, 7) trajectory of future joint targets on a fixed 10 Hz grid and executes the first 8 (n_action_steps: 8) before replanning — roughly 0.8 s per prediction, versus the ~1.1 s continuous, rescalable curve the B-spline variant emits.

Observations

key	shape	notes
top_image	(3, 128, 128)	RGB, resized from 640×480 (plain squash, no crop)
wrist_image	(3, 128, 128)	RGB, right wrist camera
joint_pos	(7,)	measured joints + gripper

Two observation steps. Images normalized to [0,1], random crop to 116×116 in training.

Two naming constraints in this codebase: the lowdim key must be named joint_pos (get_normalizer only accepts keys containing pos/quat/qpos), and the task config must set abs_action: True. At 7D→7D abs_action is a no-op for the data, but it also selects the action normalizer — False installs an identity normalizer, and joints 2/3 (up to 2.77 rad) then fall outside the sampler's clip_sample range of [-1, 1], silently breaking training.

Files

file	size	use
deploy_ema.ckpt	426 MB	Inference. EMA weights only.
epoch0600_full.ckpt	1.5 GB	model + ema_model + optimizer, for resuming or fine-tuning.

Both embed the Hydra config (pickled with dill); diffusion_policy must be importable on load.

Training

base	B-spline-policy/bspline-policy UNet DP
data	50 episodes, 37,986 frames @ 30 Hz → resampled to 10 Hz (12,677 steps)
hardware	1× RTX 4090, 2 h 00 m, ~16 it/s
epochs / batch	601 / 64
optimizer	AdamW, lr 1e-4, cosine, 500 warmup steps, EMA
params	66.9M diffusion + 22.4M vision (ResNet18 ×2)
scheduler	DDIM, 100 train timesteps, 16 inference steps, epsilon prediction
horizon / obs / action steps	16 / 2 / 8
final train_loss	0.001

Measured behavior

Over 200 held-out samples, predicted chunks vs the recorded demos:

metric	this model	B-spline variant
open-loop arm error	median 0.48°, p90 1.20°	median 0.81°, p90 1.68°
gripper predicted range	[−0.012, 1.000]	[−0.31, 1.31]
peak arm velocity	46 °/s p95, 110 °/s max	48 °/s p95, 112 °/s max
inference latency	44 ms (RTX 4090)	46 ms

Velocity was sampled at this model's native 10 Hz action rate versus 100 Hz for the B-spline curve, so those two rows are not strictly equivalent — the 10 Hz figure cannot see sub-100 ms spikes.

Because DP predicts the trajectory directly rather than control points, its outputs stay inside the demonstrated range, so it does not need the gripper clamp the B-spline variant requires.

Deployment

Environment setup, camera configuration, arm bring-up, and the five repo patches for joint-space actions are identical to the B-spline model card — follow steps 1–4 there. Those patches are applied and verified in the working tree these models were trained from, but are not in upstream B-spline-policy/bspline-policy; apply them yourself if you start from upstream. The gripper clamp in decode_action_vector is unnecessary for this model (see below) but harmless and worth keeping.

They are needed because the upstream rollout path decodes end-effector actions (its iPhone teleop records EE poses), while this model predicts joints — a difference in what was recorded, not in the LeRobot file format. Two things carry over unchanged: a missing top_image does not raise (policy_local_bspline.py:625-629 feeds a black frame, so a dead camera yields a half-blind policy rather than an error), and image resolution needs no change because the rollout resizes from the checkpoint's shape_meta. The top camera is an Intel RealSense; use TOP_CAMERA_TYPE = 'realsense' so the RGB pipeline matches training.

The gripper-convention check in the B-spline card applies here too — verify it before the first rollout, or the policy will open to grasp and close to release.

hf download Dimios45/yam-duster-dp deploy_ema.ckpt --local-dir ./ckpt

sudo ip link set can_follower_r up type can bitrate 1000000
python real_env/yam_teleop/yam_server.py --channel can_follower_r

python real_env/yam_teleop/rollout_local_policy.py \
  --env yam --policy dp \
  --ckpt-path ./ckpt/deploy_ema.ckpt \
  --diffusion-policy-dir $PWD/diffusion_policy \
  --control-freq 100 --data-freq 10 \
  --save --output-dir data/local_policy_rollouts

Note --policy dp (not bspline), and that --origin-time-scale / --predict-before-end / --speed-up-times do not apply — this model emits a fixed-rate action grid with no temporal rescaling. That is precisely the capability the B-spline variant adds.

Reproducing

python tools/lerobot_v3_to_robomimic.py \
  --repo-id Dimios45/yam_duster_in_box \
  --output diffusion_policy/data/yam_duster_in_box.hdf5 \
  --target-fps 10 --image-size 128

cd diffusion_policy && python train.py \
  --config-name=yam_duster_unet_dp \
  hydra.run.dir=../outputs/yam_duster_dp_full \
  training.resume=false logging.mode=offline \
  checkpoint.topk.k=601 dataloader.persistent_workers=True

Citation

@inproceedings{chi2023diffusionpolicy,
  title={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
  author={Chi, Cheng and Feng, Siyuan and Du, Yilun and Xu, Zhenjia and
          Cousineau, Eric and Burchfiel, Benjamin and Song, Shuran},
  booktitle={Proceedings of Robotics: Science and Systems (RSS)},
  year={2023}
}