Franka SkillGen Preflight

Complete this preflight before annotating demonstrations or starting a long SkillGen generation run. It verifies the cuRobo image, GPU build, source dataset, Nucleus asset download, and one complete end-to-end generation run. Repeat it after changing GPUs or rebuilding the development image.

This page assumes the base AutoData installation is complete and the host satisfies the support matrix.

Requirements at a Glance

Franka SkillGen requirements

Requirement	What is needed
Container	The cuRobo development image started with ./docker/run_docker.sh -c.
GPU memory	At least 24 GB VRAM is recommended for 1–2 environments; use a 48 GB GPU for approximately 5 environments.
GPU build	cuRobo kernels compiled for the compute capability of the GPU running generation.
Network	Outbound access to the Nucleus asset server during planner initialization so AutoData can retrieve the Franka URDF.
Source data	Git LFS objects pulled and a structurally valid SkillGen-annotated HDF5 dataset.
Display	Optional. Use --viz none for headless generation. A display is required for manual annotation, Kit visualization, and the Rerun plan viewer.

1. Start the cuRobo Container

On the host, confirm which GPU will run SkillGen and start the cuRobo image:

nvidia-smi --query-gpu=name,compute_cap,memory.total --format=csv
git lfs pull
./docker/run_docker.sh -c

The first build compiles cuRobo for the compute capability reported by nvidia-smi and can take considerably longer than starting the base container. If the image was built for a different GPU, rebuild it from the host:

./docker/run_docker.sh -c -r

Use -R instead of -r only when a cached build remains incompatible. See cuRobo Build and GPU Compatibility for missing-module and CUDA-kernel errors.

2. Verify cuRobo and the Source Dataset

Run these checks inside the cuRobo container:

python -c "import curobo, torch; print(curobo.__file__); print(torch.cuda.get_device_name(0))"
python scripts/validate_dataset.py ./datasets/annotated_datasets/dataset_franka_skillgen_annotated.hdf5

The first command must print the installed cuRobo path and expected GPU. The validator must report the source dataset as VALID. If it reports file signature not found or the file is an LFS pointer, follow Git LFS Pointer Files.

3. Run One End-to-End Planning Check

Still inside the cuRobo container, generate one successful cube-stacking demonstration:

python scripts/generate_dataset.py \
    --env_name Isaac-Stack-Cube-Franka-IK-Rel-v0 \
    --alg skillgen \
    --task_descriptor autodata_examples/tasks/franka_cube_stack_skillgen.yaml \
    --embodiment autodata_examples/embodiments/franka_ik_rel_skillgen.yaml \
    --input_file ./datasets/annotated_datasets/dataset_franka_skillgen_annotated.hdf5 \
    --output_file ./datasets/generated_dataset_skillgen_preflight.hdf5 \
    --result_file ./datasets/generated_dataset_skillgen_preflight.json \
    --generation_num_trials 1 \
    --num_envs 1 \
    --viz none

This check exercises the Nucleus URDF download, cuRobo initialization, collision-world update, motion planning, execution, and HDF5 recording. Validate the result:

python scripts/validate_dataset.py ./datasets/generated_dataset_skillgen_preflight.hdf5

The preflight passes when generation records one successful demonstration and the validator reports the output as VALID. A Nucleus download failure is a setup problem; follow Nucleus Asset Download before investigating planner configuration.

Expected Planning Failures

SkillGen plans against a newly randomized scene on every attempt. Some configurations have no path within the planner’s search budget, so an individual planning failure is expected. Both shipped SkillGen descriptors set guarantee_success: true. AutoData counts the failed attempt, resets the scene, and retries until it records the requested number of successful demonstrations.

Interpreting SkillGen failures

Observation	Expected?	Response
An occasional attempt reports no collision-free path, then generation continues.	Yes	No action is required. Monitor the overall success rate.
Bin stacking rejects more attempts and runs longer than plain cube stacking.	Yes	Narrow bin clearances make planning harder.
Every attempt fails planning from the same subtask.	No	Check its start annotation, end-effector frame and offset, planner profile, and collision world. See SkillGen Planning Failures.
CUDA reports out-of-memory while planners initialize or run.	No	Reduce --num_envs. Validate with one environment before scaling.
cuRobo import, CUDA kernel, or Nucleus download fails before planning begins.	No	Treat it as a preflight failure and use the corresponding troubleshooting section.