Diffusion Policy — bluev2 rs500, 50k steps
- Dataset: Kaz55/dg5f_ur5e_bluev2_rs500 — 90 episodes / 101,406 frames
- Cameras: RealSense x2 and GelSight x2, all at 500x375
- Policy: Diffusion Policy,
horizon=64,n_action_steps=60 - Training: 050000 of 200,000 steps, batch 8, seed 1000 (~15.8 epochs at 200k)
Why this dataset
LeRobot's Diffusion Policy requires every camera to share one resolution
(configuration_diffusion.py, validate_features). The regular bluev2 set has
RealSense at 640x480 and GelSight at 500x375 and is rejected outright. rs500 is
the variant with RealSense resized to 500x375 so all four cameras match.
Why horizon 64 and not 60
horizon must be a multiple of 8 — the U-Net downsamples by 2 three times — so a
chunk of 60 cannot be expressed. 64 is the nearest valid value at or above 60;
n_action_steps stays at 60 to line up with the ACT ac60 runs.
Inputs
observation.state (26) + RealSense x2 + GelSight x2.
observation.velocity / observation.effort are deliberately excluded, matching
every other run in these sweeps.
Checkpoints
| steps | model |
|---|---|
| 50k | dp-bluev2-rs500-h64-50k |
| 100k | dp-bluev2-rs500-h64-100k |
| 150k | dp-bluev2-rs500-h64-150k |
| 200k | dp-bluev2-rs500-h64-200k |
Reading the loss
Diffusion loss is the MSE of predicted noise; ACT loss is an L1 error on actions. They are not comparable. Compare diffusion checkpoints only against each other, and choose a policy by on-robot evaluation.
- Downloads last month
- 13