ywu67/b1k-Evo1-Stage1

Evo1 R1Pro Stage 1 Checkpoints

Stage 1 training checkpoints for Evo-1 model on R1Pro dataset (2025 Challenge Task 0006).

Training Details

Dataset: R1Pro 2025-challenge-demos-task0006

• 200 episodes, 1.5M frames
• 3 RGB camera views (head, left_wrist, right_wrist)
• 24-dim state features (selected from 256-dim observation)

Model Configuration:

• Base model: OpenGVLab/InternVL3-1B
• Action head: Flow Matching
• Image size: 448x448
• Horizon: 50
• Action dim: 24
• State dim: 24

Stage 1 Training (Action Head Only):

• Max steps: 5000
• Batch size: 16
• Learning rate: 1e-5
• Dropout: 0.2
• Weight decay: 1e-3
• Warmup steps: 1000
• Grad clip norm: 1.0
• Trainable parameters: 122.03M (action head + integration module)
• Frozen parameters: 648.69M (VLM)

Training Speed:

• ~3.43 seconds/step with 4 DataLoader workers + GPU TorchCodec
• Total training time: ~4.8 hours

Technology Stack:

• PyTorch 2.8.0 (CUDA 12.9)
• TorchCodec 0.8.1 (GPU-accelerated video decoding)
• DeepSpeed ZeRO Stage 2
• 4 DataLoader workers with spawn multiprocessing

Checkpoints

All checkpoints saved every 500 steps:

• step_500 through step_5000 (regular checkpoints)
• step_best - Best performing checkpoint
• step_final - Final checkpoint

Each checkpoint contains:

• mp_rank_00_model_states.pt - Model weights (~2.7GB)
• bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt - Optimizer states (~1.4GB)
• config.json - Model configuration
• norm_stats.json - Normalization statistics

Usage

For Stage 2 training (full VLM finetuning), use:

--resume --resume_pretrain --resume_path step_best

State Indices

Selected 24 state features from 256-dim observation:

• Left arm joints: 158-164 (7 dims)
• Right arm joints: 197-203 (7 dims)
• Left gripper: 193-194 (2 dims)
• Right gripper: 232-233 (2 dims)
• Trunk position: 236-239 (4 dims)
• Trunk velocity: 240-241 (2 dims)

Training Loss

Initial: 0.6719 → Final: ~0.53 (decreasing trend)