IXDLI/parallel-lerobot-dp-seed100000

dp / published LeRobot PushT protocol

Final current weights, trained from scratch for 200,000 updates, seed 100000, physical batch64. Adam1e-4, gradient clip10, cosine500, ImageNet visual normalization, all206 demonstrations. No EMA is used, matching the published LeRobot training code. Weights are not selected by test scores.

• DP calls the unmodified official LeRobot DiffusionPolicy at the code commit used by the published PushT artifact.
• All206 official v2.0 demonstrations; source videos decoded to a uint8 cache without replacing them with the Stanford archive.
• All policies use official LeRobot RGB encoder and normalizers, Adam/clip10,200000updates,batch64,seed100000, no EMA.
• Native gym-pusht reset(seed), no legacy RandomState reset override and no extra physical-action clipping.
• DP/DiT use DDPM100; TDP uses DDPM100 initialization plus reactive streaming; GRN retains8 categorical rewrites; RAIN retains its flow-matching refinement and routing.
• RAIN curriculum shares the200000-update total. Risk-head fitting holds out four trajectories from the head only; its backbone has seen all206 demonstrations.
• RAIN proposal-stage evaluation uses fixed0. Its flow refiners start with zero output, preventing an untrained refiner from corrupting proposals.
• The completed final current weights are published. Official released weights are from175k and are used only as a separate environment/loader sanity reference.
• Historical dependency and physics versions are not fully archived; protocol alignment does not assert bitwise historical reproduction.

Official code: https://github.com/huggingface/lerobot/tree/3c0a209f9fac4d2a57617e686a7f2a2309144ba2 Reference model: https://huggingface.co/lerobot/diffusion_pusht W&B: https://wandb.ai/yuyuan-wu-universiteit-gent/parallel%20lerobot/runs/parallel-lerobot-20261020-s100000-dp

Inference

Extract inference_source.tar.gz, install requirements.txt with a compatible CUDA torch/torchvision pair, then add the extracted directory to PYTHONPATH:

from pusht_lerobot.inference import load_policy, PushTController
model, payload = load_policy('policy.pt', device='cuda')
controller = PushTController(model, payload['policy'], device='cuda')
controller.reset()
action_xy = controller(observation)  # pixels uint8[96,96,3], agent_pos float[2]

DP additionally includes the official LeRobot save_pretrained format under pretrained_model/. The matching historical LeRobot source is included with its license. Metrics distinguish binary success from maximum-reward score.