Exporting to ONNX / TensorRT

Two scripts cover the export path. Both are pure-PyTorch (no Isaac Lab required).

ONNX / JIT

Specialist (residual + base)

python3 onnx_conversion.py \
    -b <path/to/x_mobility_ckpt> \
    -r <path/to/residual_policy_ckpt> \
    -o <path/to/output_onnx_file> \
    -j <path/to/output_jit_file>

Generalist

python3 onnx_conversion.py \
    -b <path/to/x_mobility_ckpt> \
    -g <path/to/generalist_policy_ckpt> \
    -e <embodiment_type> \
    -o <path/to/output_onnx_file> \
    -j <path/to/output_jit_file>

-e selects which embodiment-specific head to export. Pass one of {h1, carter, spot, g1, digit} — must match an entry in EmbodimentEnvCfgMap in run.py.

TensorRT

python3 trt_conversion.py \
    -o <path/to/onnx_file> \
    -t <path/to/trt_engine_file>

Generates a TensorRT engine specialised for your GPU. The engine is then consumed by the ROS2 deployment node.

Note

Engine portability TensorRT engines are GPU-specific. An engine built on an L40 won’t run optimally (or at all) on a Jetson Orin. Build the engine on the same GPU family as the deployment target, or use the ROS2 deployment container’s runtime conversion flow.