ACT — newblue, GelSight 88x66
One point of a GelSight-resolution sweep on the DG-5F + UR5e newblue task. Runs differ only in GelSight resolution, so any gap between them is attributable to tactile resolution alone.
- Dataset: Kaz55/dg5f_ur5e_newblue_gs88 — 90 episodes / 105,193 frames
- GelSight: 88x66
- RealSense: 640x480 (identical across the sweep)
- Policy: ACT, chunk_size=60, n_action_steps=60
- Training: 100,000 steps (~7.6 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.
Sweep
| GelSight | model |
|---|---|
| 500x375 | act-newblue-gs500-ac60 |
| 320x240 | act-newblue-gs320-ac60 |
| 160x120 | act-newblue-gs160-ac60 |
| 88x66 | act-newblue-gs88-ac60 |
| none | act-newblue-gs0-ac60 |
Caveat
On the earlier combined and blue_180ep sweeps, training loss was essentially
unchanged at every GelSight resolution, including no-GelSight-at-all. Treat
these losses as a sanity check, not as evidence about tactile resolution — that
question needs on-robot evaluation.
- Downloads last month
- 12