---
tags:
- robotics
- reinforcement-learning
- onnx
- unitree-g1
- isaac-sim
- mujoco
pipeline_tag: reinforcement-learning
---

# SONIC Capoeira V8

SONIC Capoeira V8 is a Unitree G1 whole-body control policy fine-tuned from
NVIDIA SONIC for a dynamic, hand-supported capoeira kick with landing recovery.

## Contents

- `model_step_000025_encoder.onnx`: observation encoder
- `model_step_000025_decoder.onnx`: action decoder
- `observation_config.yaml`: deployment observation configuration
- `parity.json`: PyTorch-to-ONNX numerical parity report

The paired encoder and decoder must be used together with the supplied
observation configuration and the matching motion reference dataset.

## Training and validation

The policy was fine-tuned in Isaac Sim with landing-contact, capture-point,
two-foot-support, joint-speed, knee-torque-reserve, and final-upright rewards.
It completed 112/112 Isaac Sim evaluations across nominal and randomized
profiles. ONNX parity maximum absolute error was 0 for the encoder and
1.6689300537109375e-6 for the decoder.

In the official C++ MuJoCo closed loop, the selected checkpoint passed the
initial safety gate and executed the full 467-frame motion in 10/10 trials.
Strict final post-landing stabilization succeeded in 3/10 trials.

## Safety limitation

This is a simulation research artifact. It is **not real-robot ready** and must
not be deployed on a physical G1 without additional safety validation,
hardware limits, supervision, and a reliable recovery controller.

## Code and results

https://github.com/danniely/ultimate-bots-G1

## Dataset

https://huggingface.co/datasets/hyunho7979/sonic-capoeira-v8