I4H Policy Fine-Tuning
Fine-tune GR00T or openpi PI0 policies on recorded LeRobot demonstrations.
The skill scopes work to a local dataset and GPU and does not request credentials or explicitly export data. However, it clones a remote repository, creates directories and a symlink, and starts training without explicit confirmation; permission boundaries, rollback, and data-flow disclosure are incomplete, so points are deducted.
The workflow includes environment mappings, preflight checks, verification criteria, and common troubleshooting. However, it depends on an external repository, uv, GPUs, and assumed tmux variable persistence, with no executable tests for this skill's key paths. Static calibration therefore limits the score below 10.
The audience, trigger wording, supported environments, and non-training boundaries are reasonably clear. Coverage is limited to four environments, input boundaries are incomplete, and the default GitHub clone and overseas toolchain may be unreachable from mainland China; Chinese-language support is also unspecified, so points are deducted.
The material provides a name, version, Apache-2.0 license, author, layered procedure, limitations, troubleshooting, and final reporting requirements. It lacks the recommended Instructions and Examples sections, has limited changelog, maintenance, update-path, and parameter-stability information, and repository license metadata is NOASSERTION, so points are deducted.
The skill can directly guide dataset preflight, training, checkpoint inspection, and reporting of key results, and the benchmark reports some benefit on two positive tasks. However, the benchmark has only two single-attempt tasks and lacks representative usable outputs and failure-case evidence; the static score is conservatively limited.
Versioned documentation, an evaluation summary, and internal evaluation tasks provide limited auditability. There are no raw logs, independent reproduction artifacts, negative-trigger cases, or committed tests covering this skill's key paths, so evidence remains thin.
- The default flow may clone remote code and launch GPU training without explicit user confirmation; confirm repository provenance, network access, disk space, GPU resources, and output paths first.
- Fine-tuning healthcare-robotics policies can involve sensitive data and real-world safety risks; the document lacks guidance on de-identification, isolation, model safety evaluation, and recovery after failed training.
- The benchmark covers only two positive tasks with one attempt each and does not establish boundary-trigger accuracy, abnormal-input handling, or cross-environment stability.
What does this skill do, and when should you use it?
This skill guides Isaac for Healthcare workflows for fine-tuning a policy on an existing LeRobot dataset of demonstrations. It supports GR00T and openpi PI0, selecting a training stack and CLI for the chosen environment. The workflow resolves the base code, checks the dataset and GPUs, runs training, and validates the resulting checkpoint and logs. It is for training rather than checkpoint evaluation, and inference-only environments cannot be fine-tuned.
Resolves or clones the i4h-workflows base repository; reads meta/info.json from a LeRobot dataset; checks visible GPU count and available environments; maps each supported environment to the gr00t_n15, gr00t_n16, or openpi_pi0 stack and training CLI; runs training through uv; writes checkpoints and logs under workflows/agentic/runs; and checks for safetensors files, experiment_cfg, processor, train_loss entries, and the final training summary.
- A robotics researcher needs to fine-tune a GR00T N1.5 policy on scissor_pick_and_place demonstrations.
- An embodied-AI developer needs to train a GR00T N1.6 policy for locomanip_tray_pick_and_place or locomanip_push_cart.
- A healthcare robotics team needs to train an openpi PI0 policy for ultrasound_liver_scan.
- A user with a small dataset wants to freeze the vision backbone to reduce memory use and training time.
What are this skill's strengths and limitations?
- Provides explicit environment-to-stack-to-CLI mappings.
- Includes preflight and post-run checks for datasets, GPUs, training support, checkpoints, and logs.
- Exposes practical controls for batch size, learning rate, training steps, GPU count, and vision-backbone freezing.
- Documents common setup, dataset, environment, and CLI-flag failures.
- Requires GPU hardware, with --num-gpus limited by visible GPUs.
- Requires an existing LeRobot directory containing meta/info.json.
- Environments with null policy.train_module, including assemble_trocar, cannot be fine-tuned.
- The source provides no evidence about training quality, hardware coverage, or a standalone test suite.
How do you install this skill?
Install the NVIDIA skills collection with: npx skills add nvidia/skills. Install only this skill with: npx skills add nvidia/skills --skill i4h-workflow-finetune --yes. The README does not document a separate installation directory or additional skill-specific installation procedure.
How do you use this skill?
Ask a loaded agent to fine-tune a policy on recorded LeRobot demonstrations. Prepare the i4h-workflows setup, a LeRobot dataset containing meta/info.json, and at least one visible NVIDIA GPU. Set DATASET_PATH, then run the documented preflight and training commands in order. Supported environments are scissor_pick_and_place, locomanip_tray_pick_and_place, locomanip_push_cart, and ultrasound_liver_scan.
How does this skill compare with similar options?
The skill explicitly distinguishes fine-tuning and training from checkpoint evaluation; evaluation should use i4h-workflow-validate.