ACT — blue 90ep, GelSight 500x375 (native)
One point of a GelSight-resolution sweep on the DG-5F + UR5e blue-cable task. Runs differ only in GelSight resolution, so any gap between them is attributable to tactile resolution alone.
- Dataset: Kaz55/dg5f_ur5e_blue_90ep_raw — 90 episodes / 98,812 frames
- GelSight: 500x375 (native)
- RealSense: 640x480 (identical across the sweep)
- Policy: ACT, chunk_size=60, n_action_steps=60
- Training: 100,000 steps (~8.1 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, adding a second difference between runs and breaking the ablation.
Sweep
| GelSight | model |
|---|---|
| 500x375 | act-blue-90ep-raw-chunk60 |
| 320x240 | act-blue-90ep-gs320-chunk60 |
| 160x120 | act-blue-90ep-gs160-chunk60 |
Note
On the related 180-episode combined sweep, final training loss was 0.118 across
every resolution including no-GelSight-at-all — i.e. training loss did not
detect the tactile input. Treat these numbers as a sanity check, not as evidence
that tactile resolution matters; that needs on-robot evaluation.
- Downloads last month
- -