ACT — GelSight resolution ablation (120x90)
One point of a GelSight-resolution ablation on the DG-5F + UR5e cable-sorting task. Every run in the sweep is identical apart from the GelSight resolution, so any difference between them is attributable to tactile resolution alone.
- Dataset: Kaz55/dg5f_ur5e_combined_gs120 — 180 episodes / 208,933 frames
- GelSight resolution: 120x90
- RealSense: 640x480 (unchanged across the whole sweep)
- Policy: ACT, chunk_size=60, n_action_steps=60
- Training: 200,000 steps, 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, matching the 500x375 baseline
Kaz55/act-newcable-combined-4cam-chunk60.
Letting them in via feature auto-derivation would add a second difference between
runs and break the ablation.
Sweep
| resolution | model |
|---|---|
| 500x375 | act-newcable-combined-4cam-chunk60 |
| 320x240 | act-combined-gs320-4cam-chunk60 |
| 200x150 | act-combined-gs200-4cam-chunk60 |
| 120x90 | act-combined-gs120-4cam-chunk60 |
| 64x48 | act-combined-gs64-4cam-chunk60 |
| 32x24 | act-combined-gs32-4cam-chunk60 |
| none | act-combined-gs0-4cam-chunk60 |
- Downloads last month
- 20