Views
No views yet
lerobot/smolvla_base
on makermods/3cam_200ep_blue_cube_orange_tray
— an SO-101 (6-DoF) pick-and-place: "pick up blue cube and place in orange tray".| key | resolution |
|---|---|
observation.images.front | 480x640 |
observation.images.wrist | 480x640 |
observation.images.top | 480x640 |
camera1/2/3) and training used --rename_map to bind
the real mounts onto them. Both were rewritten to the real names here, and the
rename_map cleared, so no --rename_map is needed at inference — feed
front/wrist/top directly. (lerobot-eval, lerobot-rollout and the async policy
server all override a checkpoint's saved rename map with their own, which defaults to
empty, so a checkpoint left on the placeholder keys fails there.)| steps | 20,000 (batch 64 → 32.8 epochs over 38,975 frames) |
| optimizer | AdamW, lr 1e-4, betas (0.9, 0.95), wd 1e-10, grad clip 10 |
| schedule | cosine, 1,000 warmup → 2.5e-6 at step 20,000 |
| final loss | 0.045 (grad norm 0.78) |
| trainable | 100M of 450M — frozen vision encoder, action expert only |
| hardware | RTX 4090, bf16 AMP, 19.9 GB, ~5 h |
checkpoints/<step>/ holds every 1,000-step checkpoint, each git-tagged by step, so any
can be loaded with --policy.pretrained_revision=<step>.