iFaz/eqm-pusht-seed3-half

EQM Policy — eqm_pusht_seed3

Trained with LeRobot.
Date: 2026-08-12 01:35
Policy type: eqm | Device: cuda

---

📦 Dataset

Parameter	Value
dataset.repo_id	lerobot/pusht

---

🏋️ Training Config

Parameter	Value
steps	70000
batch_size	8
eval_freq	0
save_freq	40000
num_workers	4
seed	3
eval.n_episodes	1
eval.batch_size	1
eval.use_async_envs	True

---

📐 Policy Architecture

Parameter	Value
enable_world_model	True
jepa_encoder_name	facebook/vjepa2-vitl-fpc64-256
freeze_jepa_encoder	True
jepa_tubelet_size	2
num_video_frames	8
predictor_depth	12
predictor_num_heads	8
predictor_mlp_ratio	4.0
num_action_tokens_per_timestep	8
world_model_loss_weight	0.05
aux_branch_tap	layer4
num_predictor_views	2
multicam_strategy	duplicate

---

🎯 Eval Config

Parameter	Value
env.type	pusht
env.task	PushT-v0
eval.n_episodes	100
eval.batch_size	4
eval.use_async_envs	False
policy.path	/kaggle/working/outputs/train/pusht_seed3/checkpoints/last/pretrained_model

---

📊 Eval Results

Metric	Value
Episodes	100
Success rate	21.0%
Avg sum reward	73.48
Avg max reward	0.71
Eval time (s)	534.2

---

Citation

@misc{cadene2024lerobot,
  author = {Cadene, Remi and Alibert, Simon and others},
  title  = {LeRobot},
  year   = {2024},
  url    = {https://github.com/huggingface/lerobot}
}