Fine-tune GR00T or openpi PI0 robotic policies on LeRobot datasets.
npx skills add https://github.com/NVIDIA/skills --skill i4h-workflow-finetune
Please help me install this Agent Skill. Skill: i4h-workflow-finetune Source: https://github.com/NVIDIA/skills/tree/main/skills/i4h-workflow-finetune Command: npx skills add https://github.com/NVIDIA/skills --skill i4h-workflow-finetune
Training a GR00T or openpi PI0 robot policy on recorded demonstrations requires knowing the correct stack, CLI, environment configuration, and flags for each Isaac for Healthcare workflow. This Skill guides an agent through the exact fine-tuning procedure, from dataset validation to checkpoint verification, avoiding misconfigured runs.
Ask your agent to fine-tune a policy for the scissor pick-and-place task on your converted LeRobot dataset for a short smoke run.
High-intent search queries and answers about installing and using this skill.
Point DATASET_PATH at a LeRobot directory containing meta/info.json, then run the environment's training CLI (e.g., i4h-agentic-gr00t-n15-train) with --dataset-path, --output-dir, and --max-steps. The Skill maps each environment to the correct stack and CLI automatically.
The ultrasound_liver_scan environment maps to the openpi_pi0 stack and the i4h-agentic-openpi-pi0-train CLI. Run it with your converted LeRobot dataset path, output directory, and desired step count from the i4h-workflows repository root.
Train support is determined by policy.train_module in the environment's YAML config; a null value means inference-only. The assemble_trocar environment is inference-only, while scissor_pick_and_place, locomanip envs, and ultrasound_liver_scan support training.
The training CLIs use Tyro, which requires kebab-case flags such as --max-steps rather than --max_steps. Rewrite any snake_case flags to kebab case and rerun the command.
The --no-tune-visual flag freezes the vision backbone so only the action head and projector are trained. This runs roughly twice as fast with half the memory and is a good default for small datasets.
At least one visible GPU is required, and --num-gpus must not exceed the count reported by nvidia-smi. Multi-GPU training is supported by setting --num-gpus to the desired number within that limit.
nebiusManage robot policy training, evaluation, and inference with LeRobot.
qcmuuFine-tune and serve OpenPI robot policies for manipulation datasets.
dongg622Configure Hygon DCU GPU passthrough for KVM virtual machines.
bmslabsGenerate .NET Core Web API controllers with CRUD, search, and authentication scaffolding.
