MODEL / CHECKPOINT

OmniXtreme

BIGAI / Unitree & collaborators

Model

General motion tracker combining the released flow-matching base policy, residual controller and forward-kinematics model.

Author source ↗Inference adapter ↗

Organization: BIGAI / Unitree & collaborators
Type: Flow
Flow-matching base policy with a residual controller and native motor-target limits.
Version: Author release · 6f2716d
author checkpoint
Robot support: Unitree G1 · 29 DoF
Author source revision: 6f2716d1b1d1132d513d86beb2193f5b18f11fcd

Checkpoint artifacts

Author release: OmniXtreme authors ↗
Artifact type: Author-released checkpoint
Original file: base_policy_trt.onnx + residual_policy.onnx + fk_trt.onnx ↗

All three original ONNX files match pinned Git blob SHA1 and local SHA256; no conversion.

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 ↗

b050772c0b49c813f457cdf21bb3e4f23a027e2e22c5279ab23423ea264007e2

fk
Download ↗

16c1c04d431a15964db422180039c2d0fb688eeb246d0bdbc041de9a25d756ef

residual
Download ↗

6174205a430b84ec51d3f15e014907504a9ad0c38575535a509bc771ebbc1d85

Inference interface

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

Inference implementation ↗Provenance ↗

Observation requirements

Observation rules ↗

Model-specific handling

Released actor uses joint state, angular velocity/gravity or orientation and reference features. Retain author normalization, history/recurrent state, action processing and declared lookahead; validate numerical independence from actual translation.

Exact checkpoint & evidence ↗

Control interval: 20 ms
Reference history: 0 ms
Reference preview: 20 ms
Output interface: joint-pd-feedforward
Configuration: policy.json

Evaluations

LAFAN1 × G1 29DoF · Ideal proprioception · no global position ↗