agibot_gam_h1to4_v3_c16_gb320_3436498

AgiBot GAM, step 11,000

Latest saved checkpoint checked on September 21, 2026, for W&B run agibot_gam_h1to4_v3_c16_gb320_3436498 (xvbt6gtt).

This checkpoint comes from continuation job3461453, which resumed the original job3436498 step6,858 state under the same W&B run. The continuation samples 50% AgiBot and 50% OXE using shared padded state/action inputs.

• Portable checkpoint: checkpoints/0011000.pt
• Completed training updates: 11,000
• History lengths: 1–4; action chunk: 16; global batch: 320
• AgiBot cameras: head, left hand, right hand
• Shared state/action width: 16
• AgiBot state: left/right joint and gripper values
• AgiBot action: EEF-relative arm actions plus base motion
• OXE state: slots0:7, zero padding7:16
• OXE actions: right-arm slots7:14, other slots masked
• Resolved configuration: config.yaml
• Normalization and preparation sidecars: stats/
• File sizes and SHA256: manifest.json

The portable checkpoint contains model, configuration and normalization state. Separate DeepSpeed optimizer shards are not included. This export is not a full optimizer-state training-resume backup.

Load with the matching 3DA_unified AgiBot/OXE checkpoint loader. This is a project-specific PyTorch checkpoint, not a Transformers AutoModel export. The resolved configuration preserves source-server paths; replace data and backbone paths when using another machine.

The stored training step, checkpoint ZIP structure and uploaded file hashes were verified. No new inference benchmark was run for this export. Source training code: ONground-Korea/3DA, revision e49b55a3.