real-marker-d2-r02-mulligan-no-cf-dp
Part of Mulligan. Browse evaluations at Policy Arena; the datasets are in the mulligan organization.
Diffusion-policy actor in LeRobot's checkpoint layout (config.json, model.safetensors, pre/post-processor normalizers). The config also carries Mulligan fields (camera crop boxes, action target and frame) that stock LeRobot DiffusionConfig rejects; loading needs the Mulligan training code.
| Property | Value |
|---|---|
| Task | real-marker-d2 |
| Model round | R2 |
| Arm | mulligan-no-cf-dp |
| Training step | 100000 |
Checkpoints
| Folder | Training run config |
|---|---|
| (root) | release/run-configs/real-marker-d2-r02-mulligan-no-cf-dp.json |
Evaluations
| Dataset | Session | Policy | Actor | Critic | N | Successes |
|---|---|---|---|---|---|---|
| real-marker-d2-r02-eval | b02 | mulligan_no_cf_r02_dp | real-marker-d2-r02-mulligan-no-cf-dp | — | 1 | 24/50 |
Data
Trained on: real-marker-d2-c00-teleop-sobol, real-marker-d2-c01-dagger-mulligan, real-marker-d2-c02-dagger-mulligan-no-cf.
Validation-only datasets: real-marker-d2-c00-teleop-validation.
Referenced by the metadata of: real-marker-d2-r02-eval.
Provenance
SHA-256 of every file is recorded in release.json. .pt files are PyTorch pickles; load them only in a trusted environment.
The run configs are in the Mulligan code repository; the recipes there retrain each checkpoint. License: mit.