Skip to main content
Stage 4 · Chapter 17 · Deploy on Jetson

17. Prepare the Dataset and Model Assets

Copy the converted dataset, the fine-tuned checkpoint, and the VLM backbone to the Jetson, and validate that everything the TensorRT exporter needs is present.

17.1 Dataset Compatibility​

Match training exactly

17.1 Dataset Compatibility

The converted dataset must preserve the camera keys, state/action dimensions, task labels, and episode boundaries used during fine-tuning.

For the SO-ARM101 that means:

  • camera keys wrist and front mapped in meta/modality.json, exactly as created in Chapter 10;
  • six state/action dimensions: single_arm 0–5, gripper 5–6;
  • the embodiment tag NEW_EMBODIMENT registered by so101_config.py.

Do not rename camera modalities after training.

17.2 Convert and Copy the Dataset​

v2.1 layout

17.2 Convert and Copy the Dataset

If you have not converted yet, run the converter on the workstation (the same command as Chapter 10):

cd ~/Isaac-GR00T
uv run --project scripts/lerobot_conversion \
python scripts/lerobot_conversion/convert_v3_to_v2.py \
--repo-id so101_vials_real \
--root $HOME/rebot-arm-dli-isaacsim/datasets

Copy the converted dataset (and its meta/modality.json) into the Jetson workspace path:

rsync -av \
~/rebot-arm-dli-isaacsim/datasets/so101_vials_real/ \
jetson:"${GR00T_DATASET_PATH}/"

Validate the layout GR00T expects:

${GR00T_DATASET_PATH}/
├── data/chunk-000/episode_000000.parquet
├── videos/chunk-000/observation.images.front/episode_000000.mp4
└── meta/
├── episodes.jsonl
├── info.json
├── modality.json
├── stats.json
└── tasks.jsonl
source "${HOME}/.config/gr00t-inference/paths.sh"
test -f "${GR00T_DATASET_PATH}/meta/info.json"
test -f "${GR00T_DATASET_PATH}/meta/modality.json"
test -f "${GR00T_DATASET_PATH}/meta/stats.json"
test -f "${GR00T_DATASET_PATH}/meta/tasks.jsonl"
find "${GR00T_DATASET_PATH}/data" -name 'episode_*.parquet' | head

17.3 Prepare the Fine-Tuned Checkpoint​

Weights only

17.3 Prepare the Fine-Tuned Checkpoint

Copy the checkpoint directory from Chapter 10 into ${GR00T_MODEL_PATH}, then verify:

test -f "${GR00T_MODEL_PATH}/config.json"
test -f "${GR00T_MODEL_PATH}/processor_config.json"
test -f "${GR00T_MODEL_PATH}/statistics.json"
test -f "${GR00T_MODEL_PATH}/model.safetensors.index.json"
ls "${GR00T_MODEL_PATH}"/model-*-of-*.safetensors

Training-state files such as optimizer.pt, scheduler.pt, and trainer_state.json are not required for inference or TensorRT export.

17.4 Prepare the VLM Backbone​

Cosmos-Reason2-2B

17.4 Prepare the VLM Backbone

GR00T 1.7 defaults to nvidia/Cosmos-Reason2-2B:

hf download nvidia/Cosmos-Reason2-2B \
--local-dir "${GR00T_BACKBONE_PATH}"

test -f "${GR00T_BACKBONE_PATH}/config.json"
test -f "${GR00T_BACKBONE_PATH}/model.safetensors" || \
test -f "${GR00T_BACKBONE_PATH}/model.safetensors.index.json"

The validated repository can also use Qwen/Qwen3-VL-2B-Instruct through GR00T_BACKBONE_PATH. Qwen3-VL is a validated alternative for this checkpoint and repository revision, not the official GR00T 1.7 default.

17.5 Asset Readiness Checklist​

Before Chapter 18

17.5 Asset Readiness Checklist

  • Dataset episodes and videos use the expected per-episode layout;
  • Camera keys, embodiment tag, state dimensions, and action dimensions match training;
  • The checkpoint contains all model shards and GR00T metadata;
  • The selected VLM backbone is pinned and available locally.

Next: Chapter 18 — Run GR00T N1.7 TensorRT Inference.

Loading Comments...