hackathon-everest-policies

Hackathon Everest Policies

Exports for the Hackathon Everest reduced-order, sensorized Unitree G1 crampon research project.

Safety notice: These are synthetic simulator outputs. They are research artifacts only, not a hardware deployment, a safety-certified controller, or evidence of real-world Everest readiness. Do not use them to control hardware without extensive independent validation, safety systems, and expert review.

Files

visible_policy_l0_v2/

• visible_policy.pt — PyTorch checkpoint for the visible-only terrain estimator and supervisor policy.
• visible_policy_jit.pt — traced TorchScript combined policy export.
• estimator_jit.pt — traced TorchScript sensor/terrain estimator export.
• metrics.json — synthetic sealed-test metrics and claim boundary.

The combined policy input is (packet_history, valid_mask, sample_age_s, deployable_context, deployable_command_gait_context). The first three tensors preserve the bilateral 19-channel sensor ABI: [B, T, 2, 19]. See the source exporter for exact output semantics.

terrain_estimator/

• terrain_estimator.joblib — reduced-order TerrainStateEstimator trained on synthetic terrain fields.
• manifest.json, metrics.json, benchmark.json — feature contract, held-out metrics, and simulator benchmark context.

Limits

• Synthetic priors are not calibrated to field SnowMicroPen or radar data.
• Uncertainty outputs are approximate, not certified probabilities.
• Bilateral replay and supervisor outputs are not whole-body G1 locomotion validation.
• No hardware safety, actuator, timing, packet-loss, thermal, or emergency-stop validation is claimed.

Source and license

Source code: https://github.com/joshuajerin/hackathon-everest

Released under the MIT License. Model artifacts inherit the project license.