joonyeongs/robolab-droid-200

RoboLab DROID demonstrations — Edge training by default

Dataset for joonyeongs: 200 successful episodes, 25 per task, 157,925 synchronized frames at 15 Hz. Tasks: ToyInBinTask, CannedFoodInBinTask, ToolsPickingDrillTask, RedItemsInBinTask, RecycleCartonsVerticalCrateTask, PlasticBottlesInSquarePailTask, BlockStackingSpecifiedOrderTask, SpoonsInPotTask.

Default: nvidia/Cosmos3-Edge-Policy-DROID, generator video/action LoRA, 32 samples/GPU × two GPUs × accumulation 2 = global batch 128. The recipe uses 4,000 optimizer steps with the original two-way FSDP parallelism. Measured throughput is 132.7 seconds/step, projecting to 147.4 hours (6.1 days) on two RTX A6000 48 GB GPUs. Nano is retained as an optional alternative.

Contents

• cosmos3_droid_demos/original: native HDF5, three 640×360 RGB videos per episode, camera calibration, robot/object states, all eight physics-substep motor targets and provenance.
• cosmos3_droid_demos/train_ready/droid_plus_lerobot_640x360_20260412/success: LeRobot v3 for Cosmos.
• cosmos3_droid_demos/train_ready/conditioning_edge: default Edge conditioning, all 281 prompt variants and the full-model parity fixture.
• cosmos3_droid_demos/train_ready/conditioning: retained Nano conditioning and parity fixture. Caches are model-specific; do not interchange them.
• training_defaults.json: model, batch, paths and recipe selection.
• cosmos3_droid_demos/EDGE_TRAINING.md: current Edge measurements and optimization report. The older REPORT.md is historical Nano/collection reporting.
• pi05_droid_lora/lerobot/robolab/droid_200: OpenPI's pinned LeRobot v2.1 format.
• Reports, data audits and measured batch selections for both trainers.

All 200 episodes are training data; there is no held-out split. Only target poses vary within validated feasible reset distributions. Cosmos uses the last absolute joint target per interval. OpenPI uses reconstructed normalized joint velocity (q_post - q_pre)/0.2, plus the recorded gripper target. These are not original velocity commands. Nineteen frames exceed nominal velocity bounds and are retained without clipping. Training feasibility was tested; fine-tuned policy success was not evaluated.

Base model checkpoints, environment installations, failed runs, and simulator assets are not included. Obtain model weights from their official sources below. Historical JSON provenance contains source-machine paths; runtime paths are set by the environment variables below.

Clone and download on another server

Authenticate with GitHub and Hugging Face as an account with access (for example gh auth login and hf auth login). Never place tokens in clone URLs or committed files. Install Git LFS and uv as prerequisites.

export TRANSFER_ROOT="$HOME/robolab-transfer"
mkdir -p "$TRANSFER_ROOT"
cd "$TRANSFER_ROOT"
GIT_LFS_SKIP_SMUDGE=1 git clone -b feature/robolab-cosmos3-droid-lora-demos https://github.com/joonyeongs/RoboLab.git
GIT_LFS_SKIP_SMUDGE=1 git clone -b feature/robolab-cosmos3-droid-lora-demos https://github.com/joonyeongs/cosmos-framework.git
GIT_LFS_SKIP_SMUDGE=1 git clone --recurse-submodules -b feature/robolab-pi05-droid-lora https://github.com/joonyeongs/openpi.git
hf download joonyeongs/robolab-droid-200 --repo-type dataset --local-dir "$TRANSFER_ROOT/data"

Git LFS objects for upstream sample/simulator assets remain at their original upstream repositories. Training on these downloaded demonstrations does not require those assets. To retrieve them for other uses:

git -C "$TRANSFER_ROOT/RoboLab" remote add upstream https://github.com/NVlabs/RoboLab.git
git -C "$TRANSFER_ROOT/RoboLab" lfs fetch upstream HEAD
git -C "$TRANSFER_ROOT/RoboLab" lfs checkout
git -C "$TRANSFER_ROOT/cosmos-framework" remote add upstream https://github.com/NVIDIA/cosmos-framework.git
git -C "$TRANSFER_ROOT/cosmos-framework" lfs fetch upstream HEAD
git -C "$TRANSFER_ROOT/cosmos-framework" lfs checkout

Cosmos Edge training (default)

Install the training environment from the cloned cosmos-framework/docs/setup.md (measured with Python 3.13 and PyTorch 2.13/cu130). Then change to $TRANSFER_ROOT/cosmos-framework:

cd "$TRANSFER_ROOT/cosmos-framework"
export TRANSFER_ROOT="$HOME/robolab-transfer"
export ROBOLAB_DEMOS="$TRANSFER_ROOT/data/cosmos3_droid_demos"
export COSMOS_DROID_MODEL=edge
export COSMOS_DROID_CHECKPOINT="$TRANSFER_ROOT/checkpoints/Cosmos3-Edge-Policy-DROID"
export COSMOS_DROID_CONDITIONING="$ROBOLAB_DEMOS/train_ready/conditioning_edge"
export WAN_VAE_PATH="$ROBOLAB_DEMOS/training_assets/wan22_native.pth"
export COSMOS_DROID_PER_DEVICE_BATCH=32
hf download joonyeongs/robolab-droid-200 --repo-type dataset \
  --local-dir "$TRANSFER_ROOT/data"
hf download nvidia/Cosmos3-Edge-Policy-DROID \
  --revision 0110e64ca982589f1a3d856cbdd7d465dccb79bc \
  --local-dir "$COSMOS_DROID_CHECKPOINT" \
  --include '*.json' '*.jinja' 'transformer/*' 'text_tokenizer/*' 'vae/*'
# The matching Edge conditioning cache is included in the dataset.
# Convert the bundled VAE once if its native file is not already present:
if [[ ! -f "$WAN_VAE_PATH" ]]; then
  COSMOS_DEVICE=cpu .venv/bin/python -m cosmos_framework.scripts.prepare_robolab_vae \
    --checkpoint "$COSMOS_DROID_CHECKPOINT" --output "$WAN_VAE_PATH"
fi
.venv/bin/python -m cosmos_framework.scripts.prepare_robolab_optimizer \
  --build-root "$ROBOLAB_DEMOS/training_assets/dependencies"
bash examples/launch_robolab_droid_lora.sh --dryrun
# Start the full run only when ready:
bash examples/launch_robolab_droid_lora.sh

The generic launcher dispatches to launch_robolab_droid_edge_lora.sh by default. Its explicit Edge counterpart remains available. The generic TOML examples/toml/sft_config/robolab_droid_lora.toml also selects Edge. Runtime paths are set by the environment variables above; historical reports contain source machine paths. Dataset selection metadata is in training_defaults.json.

To regenerate Edge conditioning, use bash examples/prepare_robolab_droid.sh. This requires loading the reasoning branch once; ordinary cached training does not. See the Edge guide for encoder parity checks, short training validation and checkpoint details. The memory-sweep validator now defaults to Edge; --model nano explicitly selects Nano. It deliberately probes 64/device first and is optional for the already selected 32 × 2 setup.

Optional Nano training (retained)

The Nano conditioning cache, data and training code are preserved. When ready to use Nano, download its matching checkpoint and switch both checkpoint and cache:

cd "$TRANSFER_ROOT/cosmos-framework"
export COSMOS_DROID_MODEL=nano
export COSMOS_DROID_CHECKPOINT="$TRANSFER_ROOT/checkpoints/Cosmos3-Nano-Policy-DROID"
export COSMOS_DROID_CONDITIONING="$ROBOLAB_DEMOS/train_ready/conditioning"
export COSMOS_DROID_PER_DEVICE_BATCH=16
hf download nvidia/Cosmos3-Nano-Policy-DROID --local-dir "$COSMOS_DROID_CHECKPOINT"
bash examples/launch_robolab_droid_nano_lora.sh --dryrun
# Start only when ready:
bash examples/launch_robolab_droid_nano_lora.sh

Use prepare_robolab_droid_nano.sh if regeneration is needed. Nano's recipe is examples/toml/sft_config/robolab_droid_nano_lora.toml; the generic shell commands can also select it with COSMOS_DROID_MODEL=nano. The shared Wan VAE weights are identical in the checked Edge and Nano releases. A clean installation on a second server has not been tested here. No full fine-tuning run was executed.

OpenPI training

cd "$TRANSFER_ROOT/openpi"
GIT_LFS_SKIP_SMUDGE=1 uv sync --python 3.11
export ROBOLAB_PI05_ROOT="$TRANSFER_ROOT/data/pi05_droid_lora"
export PI05_DROID_CHECKPOINT="$TRANSFER_ROOT/checkpoints/pi05_droid"
.venv/bin/python scripts/robolab/download_checkpoint.py
bash scripts/robolab/launch.sh 1 --dry-run
# Start training only when ready (select 1, 2, or 4):
bash scripts/robolab/launch.sh 1

The variants scale language rank/alpha from 16/16 and action expert rank/alpha from 32/32 by 1, 2, or 4. Upstream freeze policy also trains vision and action/time projections. All three were tested at 64/device × two A6000 GPUs × accumulation 1 = global batch 128. Default schedule: 20,000 optimizer steps; measured estimate about 72–73 hours. Add --resume for the same full run.