act_vision_charger_50ep

ACT policy — vision only — trained on 50 episodes.

Task: grab and remove the charger from the socket and put it in the black box

Trained with LeFlexiTac, a LeRobot fork adding FlexiTac tactile sensing. Docs: https://tna001-ai.github.io/LeFlexiTac/docs.html

Training data

dataset	aryankakad/tactile_charger_inserting
episodes	50 (all)
frames	27,973 @ 30 fps
cameras	observation.images.top, observation.images.gripper (224×224)
tactile	not used — vision-only baseline
state / action	6-DoF SO-100 follower

Configuration

steps	100,000
batch size	32
epochs	114.4
chunk_size	100
n_action_steps	100
vision_backbone	resnet18
dim_model	512
n_encoder_layers	4
n_decoder_layers	1
use_vae	True
kl_weight	10.0
optimizer_lr	1e-05
optimizer_weight_decay	0.0001

Every model in this series is epoch-matched at ~114.4 epochs, so dataset size and sensor modality are the only variables across the set.

Training command actually used

Run on 1× AMD Instinct MI300X (ROCm 6.2.4). HIP_VISIBLE_DEVICES selected the GPU, so --policy.device=cuda refers to that single card.

python -u -m lerobot.scripts.lerobot_train \
  --dataset.repo_id=aryankakad/tactile_charger_inserting \
  --policy.type=act \
  --policy.repo_id=Dimios45/act_vision_charger_50ep \
  --policy.private=true --policy.device=cuda \
  --output_dir=outputs/train/A_act_vision_50ep --job_name=A_act_vision_50ep \
  --batch_size=32 --num_workers=8 --steps=100000 --save_freq=20000 --wandb.enable=true

Evaluation / rollout

Not run here — this machine has no robot attached. To evaluate, run on the machine with the SO-100 and sensors, loading the policy with --policy.path=Dimios45/act_vision_charger_50ep.

Reference: the lerobot-record eval invocations in tactile_cmd.txt and the project docs. You will need to supply your own robot port, camera serials.

Notes

• Two ROCm-specific fixes were required in the fork: persistent_workers=True on the dataloader (epoch boundaries otherwise stalled ~410 s each), and keeping cudnn.benchmark off (on ROCm it triggers an exhaustive MIOpen search that can precede step 1 by hours).
• Training loss is not a proxy for task success. Compare policies by rollout success rate, especially on contact-rich phases.
• The source dataset's task string is labelled stack cup — a mislabel carried over from an earlier session. It does not affect ACT, which is not language-conditioned.