diff --git a/src/lerobot/policies/rlearn/configuration_rlearn.py b/src/lerobot/policies/rlearn/configuration_rlearn.py index 9a1d41f17..64dc974e0 100644 --- a/src/lerobot/policies/rlearn/configuration_rlearn.py +++ b/src/lerobot/policies/rlearn/configuration_rlearn.py @@ -61,7 +61,7 @@ class RLearNConfig(PreTrainedConfig): use_tanh_head: bool = False # when True, bound outputs in [-1, 1] # Training - learning_rate: float = 3e-5 + learning_rate: float = 1e-4 weight_decay: float = 0.01 # Performance optimizations diff --git a/src/lerobot/policies/rlearn/rlearn_plan.md b/src/lerobot/policies/rlearn/rlearn_plan.md index 9bd4bb4d4..66b307a76 100644 --- a/src/lerobot/policies/rlearn/rlearn_plan.md +++ b/src/lerobot/policies/rlearn/rlearn_plan.md @@ -82,15 +82,17 @@ _ Open X-Embodiment (OXE) - benchmark siglip 2 vs this implementation forward pass, debug speed [x] - use siglip 2 [x] - Overfit on one episode [] -- Cleanup code? [] +- Cleanup code? [] + enable language loss - Convert python -m lerobot.datasets.v21.convert_dataset_v20_to_v21 --repo-id=IPEC-COMMUNITY/bc_z_lerobot and train on 1 percent -- Then on 10 percent -- Ablation dino v2 vs dino v3 base 86 M +- Then on 10 percent [] +- Ablation 16 sucessive frame vs 16 frame samples with stride 2 or 4 [] - Add more artificial text to dataset generated by vlm (google gemini) [] - See google gemini vlm caption [] https://gemini.google.com/app/7e332ffaf32580f2 - Multiple captions per video, creat method to generate as much data as possible etc [] https://arxiv.org/abs/2508.13446, https://arxiv.org/pdf/2412.04453 -- How can we improve spatial aware learning? solve issue of Contrastive learning and position -- Extend evaluation [] - Add other datasets from OXE metioned in rewind [] +- Extend evaluation [] - Ablation for size vision encoder, language encoder, temporal head [] - Add other datasets metnioned here [] +- How can we improve spatial aware learning? solve issue of Contrastive learning and position [] + + diff --git a/src/lerobot/scripts/train.py b/src/lerobot/scripts/train.py index 6594e81b5..60f66379a 100644 --- a/src/lerobot/scripts/train.py +++ b/src/lerobot/scripts/train.py @@ -258,7 +258,7 @@ def train(cfg: TrainPipelineConfig): pin_memory=device.type == "cuda", drop_last=False, persistent_workers=cfg.num_workers > 0, # Keep workers alive between epochs - prefetch_factor=12, # Maximum prefetch for video pipeline + prefetch_factor=6, # Maximum prefetch for video pipeline timeout=30, # Prevent hanging on video decode errors ) dl_iter = cycle(dataloader)