mcMotionChallengeGitHub ↗

MODEL / CHECKPOINT

RoboParty

Model

The released G1 forward-backward controller from UFO. A backward encoder turns reference poses into a latent command for its history-conditioned policy.

Organization
RoboParty
Type
FB
Forward-backward latent policy with a backward reference encoder (FB-CPR-Aux).
Version
FB-CPR-Aux · 191,995,904 training steps
author-released checkpoint; 192M denotes training steps
Robot support
Unitree G1 · 29 DoF
Author source revision
ee5c57082bb4223bdac215bff7a607cbe0a93cd1

Checkpoint artifacts

Author release
RoboParty / UFO ↗
Artifact type
Author-released checkpoint
Original file
FBcprAuxModel.onnx ↗

Official project links the artifact repository; its release manifest binds the training and deployment commits and the SHA256 of both ONNX files.

Download links deliver the adapter files. The release links and verification notes distinguish author files from conversions; SHA256 identifies the bytes used.

Policy checkpoint
Download ↗
fe58bf17427679a0f0a22ad5f05d7579373a365936154770f31e815e8f0e4ad7
encoder
Download ↗
4333441ba4a393f9493545f5c8a706fa912e8ec358dae1c05c09e31aaae00d4b

Inference interface

The adapter declares its control clock and reference window below. Source and configuration describe the observation and action transforms.

Observation requirements

Observation rules ↗

Model-specific handling

The encoder input named privileged_state is computed exclusively from REFERENCE frames. Actual actor inputs are proprioception/history; released slew limits and endpoint behavior are retained.

Exact checkpoint & evidence ↗
Control interval
20 ms
Reference history
0 ms
Reference preview
80 ms
Output interface
joint-pd-feedforward
Configuration
policy.json

Evaluations