feat: add MultiLeRobotDataset with weighted sampling and RoboMME env integration

Multi-dataset training support:
- NewMultiLeRobotDataset with per-dataset feature mapping, auto-padding,
  per-dataset transform pipelines, and weighted sampling
- MultiDatasetMeta shim compatible with EpisodeAwareSampler and make_policy
- WeightedEpisodeAwareSampler for proportional cross-dataset sampling
- SubDatasetConfig / MultiDatasetConfig in training configs
- DatasetTransformPipeline with built-in PadAction, PadState, ResizeImages
- Factory and training script wired up for multi-dataset path

RoboMME environment integration:
- RoboMMEEnv config and Gymnasium wrapper (robomme.py)
- robomme optional dependency in pyproject.toml

Made-with: Cursor
This commit is contained in:
pepijn
2026-03-13 04:31:35 +00:00
parent db5c26f07d
commit b4d40d0228
11 changed files with 886 additions and 67 deletions
+3
View File
@@ -180,6 +180,9 @@ libero_plus = [
"libero @ git+https://github.com/sylvestf/LIBERO-plus.git@main ; sys_platform == 'linux'",
"lerobot[scipy-dep]",
]
robomme = [
"robomme @ git+https://github.com/RoboMME/robomme_benchmark.git@main ; sys_platform == 'linux'",
]
metaworld = ["metaworld==3.0.0", "lerobot[scipy-dep]"]
# All