hebbianrobotics/g1-depth-dodge

G1 Depth Dodge (Hebbian Robotics)

This is the checkpoint labelled π₂ in the paper, one of the two fixed G1 depth-dodgeball policies evaluated in Beyond Dodge Rate: Testing Humanoid Dodgeball Under Camera Failures (Kingston Kuan and Brandon Ong, Hebbian Robotics).

π₂ was trained by Hebbian Robotics with the same PPO method, architecture and objective as the released G1 Depth Dodge policy (π₁), by Martino Russi. It ran 4,000 updates with 512 environments spread across eight GPUs. All 60 recorded configuration fields match the reference, including seed 42. The full training history is not available, so differences between π₁ and π₂ should not be attributed to any specific training difference.

• Paper and code: https://github.com/Hebbian-Robotics/beyond-dodge-rate
• Evaluation trial data: https://huggingface.co/datasets/hebbianrobotics/beyond-dodge-rate

Files

File	SHA-256
actor.onnx	22c6fd3c437558947d29874a61e35a57870346e5d03abbb1a4c8951567a43e01
perception.onnx	250c5e76f4393b8e41d63e7205211bde88c1b14e4fead670b4843e40b34d636f
manifest.json	Training configuration and I/O contract (iteration 4000)

These ONNX files were exported from training checkpoint model_4000.pt (SHA-256 23069f416d89f257c4ec889d47d9cac5f56570a2822f03e08bf9fc72f4d58714). The PyTorch checkpoint and its optimizer state are not distributed.

The frozen encoders sonic_plain.onnx and theia_image.onnx are byte-identical to those in the π₁ release. Download them from that revision, which also documents the input and output contract. actor.onnx and perception.onnx replace the corresponding π₁ files.

Intended use and limitations

This checkpoint is a fixed research artifact for reproducing the paper's simulated camera-failure evaluations. It has been evaluated only in MuJoCo simulation. It is not validated for physical robots or for safety-critical use. Under the paper's healthy-camera evaluation it succeeded in 161 of 192 trials. Its responses to stale and blank depth images differ from π₁'s; see the paper.

License and attribution

actor.onnx contains the frozen NVIDIA SONIC decoder with folded LoRA adapters, making it a derivative of SONIC. These files are distributed under the NVIDIA Open Model License; a copy is included in NVIDIA-OPEN-MODEL-LICENSE.txt.

Licensed by NVIDIA Corporation under the NVIDIA Open Model License.

The training method and reference policy are by Martino Russi (G1 Depth Dodge). The perception stack uses Theia; Theia weights are not redistributed here and retain their original license.

Citation

@misc{kuan2026beyonddodgerate,
  title  = {Beyond Dodge Rate: Testing Humanoid Dodgeball Under Camera Failures},
  author = {Kuan, Kingston and Ong, Brandon},
  year   = {2026},
  howpublished = {\url{https://github.com/Hebbian-Robotics/beyond-dodge-rate}}
}