Kaz55/act-blue-180ep-gs88-ac16

ACT — blue 186ep, GelSight 88x66, chunk 16

Part of a two-axis sweep on the DG-5F + UR5e blue-cable task: GelSight resolution x action-chunk length. Everything except those two knobs is pinned.

• Dataset: Kaz55/dg5f_ur5e_blue_180ep_gs88 — 186 episodes / 186,152 frames
• GelSight: 88x66
• RealSense: 640x480 (identical across the sweep)
• Policy: ACT, chunk_size=16, n_action_steps=16
• Training: 100,000 steps (~4.3 epochs), batch 8, seed 1000

Inputs

observation.state (26) + RealSense x2 + GelSight x2

observation.velocity and observation.effort are present in the dataset but deliberately excluded — feature auto-derivation would otherwise feed them to the policy and add a second difference between runs.

Reading the loss

Do not compare loss across chunk lengths. The ac16 and ac60 runs predict 16 and 60 future actions respectively, so the loss scale differs; ac16 numbers look lower simply because each prediction spans less time. Only same-chunk comparisons are meaningful, and the resolution question ultimately needs on-robot evaluation.

On the related 180-episode combined sweep, training loss was 0.118 at every resolution including no-GelSight-at-all — training loss did not detect the tactile input at all.