P-BFM G1 Deployment Models
Three matching ONNX networks for the P-BFM deployment code. A417 tracking, live teleoperation, reward and goal tasks share one policy; task contexts and deployment configuration are distributed with the code.
| File | Input | Output |
|---|---|---|
actor.onnx |
state 64, last_action 29, history_actor 372, heading 2, velocity_command 3, depth_latent 256, z 256 | action 29 |
depth_encoder.onnx |
float32 depth [1, 8, 18, 32] in [0, 1] |
depth_latent 256 |
backward_encoder.onnx |
retargeted mimic observations | z 256 |
Download all three files into model/g1_policy/onnx/ of the deployment repository, then run bash run.sh verify. Follow the repository README for the pinned model revision, dependencies, camera preprocessing and target-specific TensorRT build. Do not substitute the actor for a full depth-input model or mix encoders from other checkpoints.
SHA256SUMS records the file hashes. The code repository includes model/g1_policy/manifest.json, the deployment contract, task contexts and reference outputs.
Provenance and validation
The source checkpoint is P3 resume, step 364,822,528. The actor and depth encoder are split deployment exports; 32 synthetic comparisons against the original complete ONNX produced a maximum action error of 0.0. The release also passed 143 offline tests, including 256 frames of recorded simulation replay.
This repackaged release has not completed clean installation and native compilation on Jetson ARM64, target TensorRT inference, live RealSense/Pico validation, or motor/firmware timeout validation. Offline checks do not establish safe balance or physical emergency-stop behavior. See the deployment repository's validation record for the exact scope.
License
CC BY-NC 4.0. See LICENSE and the deployment repository for third-party component notices.