Overview#
First identify whether the error occurs during source checkout, building, Python loading, task selection, or policy loading, then check the dependencies for that stage. Check build artifacts and file paths before proceeding to device operation.
Build and Python extensions#
| Symptom | What to check |
|---|---|
install/setup.bash does not exist |
Whether the first build completed; this file is generated |
| ROS package or extension cannot be found | Whether the current terminal has loaded the ROS 2 and workspace environments |
ModuleNotFoundError: imu_py/motors_py/robot_py |
Build artifact paths, the active Python interpreter, and the environment |
.so import fails |
CPU architecture, Python ABI, and linked dependencies; do not mix artifacts from different images |
| SciPy is missing when running the complete IMU example | Install the example's SciPy dependency in the same Python environment |
| ONNX Runtime build fails | The archive selected for the platform, its version, and CMake errors; retain the first actual error |
See Source checkout and build for the build and environment-loading sequence. SDK modules use the workspace build artifacts.
Training tasks and assets#
If an Isaac Lab-related import cannot be found, check that the active Python interpreter belongs to the Isaac Lab environment and that robolab and rsl_rl are installed in that same environment. list_envs.py filters only RPO names. When checking RP1.3, see Task definitions.
roboparty_train contains an RPO asset submodule, which can also affect bulk task imports. Retain the specific missing path or import exception; do not replace RP1 geometry or substitute an RPO task simply because a directory is empty.
Policy resources#
| Configuration | Known missing files |
|---|---|
28-action rp1/default |
models/policy.onnx |
24-action rp1_24dof/ufo |
FBcprAuxModel.onnx, latents/zs_66.npz |
The presence of a file named policy.onnx alone does not establish that the I/O, external data, and joint mapping match. The deployment implementation checks for a single input and its element count; observation fields, history layout, normalizer, and order must still be verified against the I/O conventions.
For configuration errors, also check the counts and correspondence of model_names, obs_layouts, frame_stacks, and obs_stack_orders, as well as the range of history taps. Keep the input dimensions consistent with the weights' I/O conventions.