使用 dynamics=quad（四轴动力学模型）时导出 ONNX 报错

[bigbigworld123]

运行命令：
python script/train.py env=oa algo=shac dynamics=quad sensor=camera n_envs=1024 n_updates=1000 log_freq=10 device=0
报错信息：

Exception summary

<class 'ValueError'>: Unknown action frame: body

(Refer to the full stack trace above for more information.)

Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.

[bigbigworld123]

Also, using quad to train SHAC performs poorly. Is it because the simulator is not compatible?

[ZombieRUNs]

Thanks for your interest and for trying out the simulator.

Regarding the poor performance when training SHAC with quadrotor dynamics, this is expected at the current stage. In our obstacle avoidance environments, the pointmass dynamics are much more mature and stable, while the full quadrotor dynamics are still in an early, experimental stage.

For differentiable algorithms like SHAC, the full quadrotor dynamics introduce a very complex computation graph, which often leads to unstable or poor optimization. More generally, full quadrotor dynamics are better suited for standard reinforcement learning methods rather than differentiable approaches, regardless of the task.

For best results, we recommend using pointmass dynamics when working with differentiable methods.
If you have further questions, feel free to let us know.