| --- |
| library_name: lerobot |
| pipeline_tag: robotics |
| tags: [act, lerobot, robotics, dg5f, ur5e, gelsight] |
| --- |
| |
| # ACT — nutv4, chunk 40 |
|
|
| - **Dataset**: [Kaz55/dg5f_ur5e_nutv4](https://huggingface.co/datasets/Kaz55/dg5f_ur5e_nutv4) — 60 episodes / 83,311 frames |
| - **GelSight**: 500x375 (native) x2 |
| - **RealSense**: 640x480 x2 |
| - **Policy**: ACT, chunk_size=40, n_action_steps=40 |
| - **Training**: 100,000 steps (~9.6 epochs), batch 8, seed 1000 |
| |
| ## Inputs |
| |
| `observation.state` (26) + RealSense x2 + GelSight x2 |
| |
| `observation.velocity` and `observation.effort` exist in the dataset but are |
| **deliberately excluded** — feature auto-derivation would otherwise feed them to |
| the policy and add a second difference between runs. |
| |
| ## Chunk-length pair |
| |
| | chunk | model | final train loss | |
| |---|---|---| |
| | 40 | act-nutv4-ac40 | 0.085 | |
| | 60 | act-nutv4-ac60 | 0.085 | |
| |
| Both converged to the same training loss. Note that loss is normally not |
| comparable across chunk lengths (they predict a different number of future |
| actions), and on the related blue-cable sweeps training loss stayed flat even |
| when GelSight was removed entirely. Pick a chunk length by on-robot evaluation, |
| not by these numbers. |
| |
