seyilay/act_kl5_e5

ACT on SO-101, red cube into the round target, 5 epochs

Trained 2026-10-15. This is the final checkpoint (step 20,373) of act_kl5_e5.

The policy is ACT with a variational objective (kl_weight 5.0, latent dim 32). It was trained on a cleaned local copy of dharak22/so101-red-cap-pick-place-demo-v2: 50 episodes, 36,144 frames, one task ("red cube into the round target"). The Hub copy of that dataset has extra episode parquet files and was not used.

Training

• 20,373 optimizer steps. With batch size 8 that is 5.00 epochs on the training split.
• eval_split 0.1 holds out the last 5 episodes (3,548 frames). Training used episodes 0–44 (32,596 frames).
• Eval and a checkpoint every 4,075 steps, which is one epoch. This repo is the last checkpoint only.
• Chunk size 50, n_action_steps 15, ResNet18 backbone, learning rate 1e-5.
• Image transforms were configured and left disabled.
• One NVIDIA L40. The run took about 35 minutes and used 3.64 GB of GPU memory.

Last logged window

W&B logs every 200 steps, so these numbers are step 20,200 of 20,373 (epoch 4.96), not a separate measurement at the final step.

Metric	Value
train/l1_loss	0.095
train/kld_loss	0.0012, summed over 32 latent dims
train/loss	0.101, which is l1 + 5 × kld
eval/eval_loss	0.289

eval_loss is that same sum on the 5 held-out episodes. It is not holdout action error by itself. At this point the KL penalty is about 0.006 of the training loss, so the latent is close to the standard normal.

Not done

No robot rollout yet. That is waiting on the SO-101 setup.

from lerobot.policies.act.modeling_act import ACTPolicy

policy = ACTPolicy.from_pretrained("seyilay/act_kl5_e5")