| --- |
| library_name: lerobot |
| pipeline_tag: robotics |
| tags: [act, lerobot, robotics, dg5f, ur5e, gelsight] |
| --- |
| |
| # ACT — v6 blue89 180ep, GelSight 500x375 (native) |
|
|
| - **Dataset**: [Kaz55/dg5f_ur5e_v6_blue89_180ep](https://huggingface.co/datasets/Kaz55/dg5f_ur5e_v6_blue89_180ep) — 180 episodes / 208,109 frames |
| - **GelSight**: 500x375 (native) |
| - **RealSense**: 640x480 |
| - **Policy**: ACT, chunk_size=60, n_action_steps=60 |
| - **Training**: 100,000 steps (~3.84 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. |
| |
| ## Caveat |
| |
| This is a different dataset from the `blue_180ep` sweep (186 ep / 186,152 frames), |
| so losses are not directly comparable across the two. On the related `combined` |
| sweep, training loss was identical at every GelSight resolution **including |
| no-GelSight-at-all**, so treat loss as a sanity check rather than evidence about |
| tactile resolution; that needs on-robot evaluation. |
|
|
