zimplex/diffusion-genesis-hr-bench-qpos-abs-joint-step6250

Diffusion Policy (lerobot) — genesis-hr-bench final checkpoint (step 6,250, qpos_abs)

Final baseline checkpoint used to produce the genesis-hr-bench evaluation numbers (eval run full_20260529_0220, 45-task eval_v1_100 set). Inference-only (lerobot pretrained_model/ layout: weights + config + pre/post-processors).

Provenance

Base model	trained from scratch (no pretrained init)
Dataset	zhouqh/hrbench → genesis_hr_bench_lerobot_qpos_target_abs_20260526_1651
Action space	qpos_abs (8-D Franka joint absolute + gripper)
Step	6,250
Wandb	diffusion_genesis_hr_bench/ue30ic9m
Uploaded	2026-08-05 06:59:44 UTC

Note. This is the b64s6250_2x8 (2-node) variant — the one used in the eval. Two other diffusion variants exist locally (…_qpos_abs_joint @10k, …_b64s12500 @10k) but were NOT evaluated. Distinct from runs/dp3/ (3D Diffusion Policy, per-task).

Usage (lerobot)

from lerobot.policies.factory import make_policy_from_pretrained

policy = make_policy_from_pretrained("zimplex/diffusion-genesis-hr-bench-qpos-abs-joint-step6250", device="cuda")
policy.eval()
# Observation keys: observation.images.base_0_rgb / left_wrist_0_rgb /
# right_wrist_0_rgb (as available), observation.state. Action space: qpos_abs.
action_chunk = policy.predict_action(batch_observation)

Pair the checkpoint with the matching qpos_abs dataset variant — the normalizers and action_dim are baked in.