Mimicgen Coffee Ckpt

--- license: other tags: - robotics - robomimic - mimicgen - coffee - diffusion-policy --- # MimicGen Coffee_D2 Baseline Checkpoints ResNet34 image-policy checkpoints trained on: `/home1/ncloud/data/coffee_d2_240_success200.hdf5` Dataset metadata: - Env: `Coffee_D2` - Cameras: `agentview`, `robot0_eye_in_hand` - Image size: `240x240` - Crop: `224x224` - Horizon: `7` - n_obs_steps: `2` - n_action_steps: `6` - Training seed: `42` - Test start seed: `100000` ## Latest Eval Snapshot | Model | Latest eval epoch | Train SR | Test SR | Best observed test SR | | --- | ---: | ---: | ---: | ---: | | DP-C | 250 | 0.500 | 0.420 | 0.420 | | MaIL | 150 | 0.500 | 0.300 | 0.500 | | ManiFlow | 100 | 0.000 | 0.020 | 0.020 | | DiSPo | 50 | 0.167 | 0.200 | 0.200 | ## Notes - These runs used the patched Coffee_D2 eval reset path that restores dataset `model_file` and initial `states[0]`. - Camera override mode for this dataset is `none`. - The local eval summaries are included under `metadata/ /`. - Dataset action open-loop replay was checked separately. Stored final states are successful, but action replay under the current stack only reached official task success for `54/200` demos, indicating open-loop replay drift / stack sensitivity rather than a simple stored-state failure. ## Layout ```text checkpoints/ dpc/{latest.ckpt,best.ckpt} mail/{latest.ckpt,best.ckpt} maniflow/{latest.ckpt,best.ckpt} dispo/{latest.ckpt,best.ckpt} metadata/ /{latest_eval.txt,eval_summary.csv,eval_summary.jsonl,hydra_config.yaml} replay_success_check/{replay_success.csv,replay_success.jsonl,replay_success_summary.*} ```