BG2 Left Wrist Latent Flow Pretrain AE256

Single-camera, left-arm-only BG2 pretrain checkpoint.

Setup

  • Dataset: rllab-postech/pretrain_aiworker_bg2_lance
  • Camera: observation.images.cam_wrist_left
  • State transform: bg2_left_arm (19D -> 8D)
  • Action transform: bg2_left_arm_link_xyz (19D -> 39D FK points)
  • Vision encoder: DINOv3, frozen
  • Action autoencoder: point_temporal
  • Latent action dim: 256
  • Horizon: 12
  • Train steps: 30000

Result

  • Final train loss: 0.121042
  • Final val loss: 0.132042

Some source wrist videos are shorter than their metadata. Training used the loader-side frame clamp added in commit fa35ae3.

Files

  • checkpoints/bg2_left_wrist_dinov3_frozen_latent_ae256_flow_h12_30000step_final.pt
  • config.yaml
  • metadata.json
  • normalizer.json
  • results/run_summary.json
Downloads last month
-
Video Preview
loading