jaeikkim/fr3-cube-full10k-gaussian-policy

FR3 Cube Full10K Gaussian Policy

Checkpoints from 10K through 100K optimizer steps for the FR3 3-cube Gaussian Policy run.

Training contract

• ImageNet-pretrained ResNet18 for each of 3 RGB views
• Native 320×180 RGB, 4 observation frames, 10 Hz source data
• 7D delta EEF/OSC action: relative Cartesian 6D + binary gripper
• Teacher Gaussian mean/std KL loss
• Absolute and target-cube-frame relative pose auxiliary losses
• First episode frame masks action KL only; RGB and pose auxiliary remain enabled
• 2 GPUs, batch/GPU 352, global batch 704, LR 1.1e-3

Each .ckpt contains model and optimizer states plus global-step/epoch/LR-schedule resume metadata.

Sources

• Training code
• UWLab integration
• Training dataset
• W&B run

Use SHA256SUMS to verify downloaded files. Access is public with manual approval gating.