ACT — bluev2_cable170_330_270ep, 300k steps
Intermediate checkpoint from a 300k-step run, published every 100k so the effect of longer training can be compared directly.
- Dataset: Kaz55/dg5f_ur5e_bluev2_cable170_330_270ep
- Steps: 300000
- Policy: ACT, chunk_size=60, n_action_steps=60, batch 8, seed 1000
- GelSight: 500x375 (native) x2 — RealSense: 640x480 x2
observation.velocity and observation.effort exist in the dataset but are
deliberately excluded, matching every other run in these sweeps.
Checkpoints from this run
| steps | model |
|---|---|
| 100k | act-bluev2-cable170-330-270ep-ac60-100k |
| 200k | act-bluev2-cable170-330-270ep-ac60-200k |
| 300k | act-bluev2-cable170-330-270ep-ac60-300k |
Training loss keeps falling well past 100k on these datasets (an earlier 200k run improved ~25% from 100k to 200k), so later checkpoints are not merely redundant. Final policy choice still needs on-robot evaluation.
- Downloads last month
- 10