RoboDemo Evaluation Image
A Linux amd64 Docker image shared by all five RoboDemo checkpoints. It contains the pinned Python/CUDA/PyTorch environment and Imitator Game simulator source. Dataset assets and model weights are mounted separately. The runtime image omits build-stage downloads and CUDA compiler tools; the C compiler needed by Triton remains included.
1. Load The Image
On a Linux x86-64 GPU host with Docker and NVIDIA Container Toolkit:
curl -fL https://huggingface.co/mxk1998/RoboDemo-Docker/resolve/main/robodemo-evaluation-cu124-amd64.tar.gz \
| gzip -d | docker load
docker run --rm --gpus all robodemo-evaluation:cu124
The streaming command avoids keeping a second copy of the image archive. Docker still needs disk space for the unpacked image.
2. Mount Data And Models
Clone RoboDemo and follow its
data and encoder setup.
The shared asset directory contains IG-10K-Dataset/, maniskill/,
Qwen3-VL-4B-Instruct/, vjepa2/, and vjepa2-ac-vitg.pt.
Put a released checkpoint in checkpoints/RoboDemo-c15/pytorch_model.pt.
cd RoboDemo
export ROBODEMO_HOST_ASSETS=/data/robodemo-assets
export ROBODEMO_HOST_OUTPUTS="$PWD/outputs"
export ROBODEMO_GPU=0
mkdir -p "$ROBODEMO_HOST_OUTPUTS"
docker compose -f environments/docker-compose.yml run --rm evaluation \
python environments/check_environment.py --assets
3. Evaluate A Checkpoint
docker compose -f environments/docker-compose.yml run --rm evaluation \
bash scripts/evaluate/c15-single.sh \
--checkpoint checkpoints/RoboDemo-c15/pytorch_model.pt \
--data-root /assets/IG-10K-Dataset --workers-per-gpu 1 --max-episodes 1 \
--output /outputs/c15-smoke
Replace c15 with full50, c30, c45 or c15-wo and its checkpoint.
Remove --max-episodes for the complete trial plan.
GPU sizing
recommends one worker on 32 GB or two on 80 GB, based on H200 process-memory
measurements. The Compose example allocates one GPU.
For driver setup and diagnostics, use the environment guide. The reproducible Dockerfile is maintained in GitHub.
Validation
The image passed a clean build, CPU package checks and 100 regression tests for each policy variant. It also completed a 500-step Full50 episode on H200, with Vulkan/OptiX rendering and 21.38 GiB peak process memory. It completed 500 policy calls and 499 memory updates, with task success.
The GPU run loaded this exported image with Apptainer 1.5.5-1, NVIDIA driver 570.195.03 and the image's own Python packages. Docker Engine/NVIDIA Container Toolkit execution was not tested on that host. Reports: CPU environment, CPU tests and GPU validation.
The two minimal examples also passed inside the image: human-video action prediction on H200 and raw-media reference retrieval on CPU. The GPU report includes both outputs.
The host supplies its NVIDIA driver, NVIDIA Container Toolkit, Vulkan discovery and matching OptiX runtime. Do not install a kernel driver inside the image.
If the host exposes OptiX separately, keep libnvoptix.so.1 and nvoptix.bin
together and use the documented
OptiX mount override.
Licenses
Bundled CUDA, Python, simulator and dependency packages retain their upstream licenses. RoboDemo code is distributed separately under the MIT license.