g05-docker

G05 newdata step2000 — IROS 2026 Docker image

This image is built on the official challenge_base:20260806 image and keeps its CUDA 12.2, ROS 2 Humble, Redis/tmux control flow, working directory, and NVIDIA entrypoint.

Image tag: g05-newdata-step2000:iros2026-fa2-fmonly-pathfix

Checkpoint inside the image:

/opt/g05/run/checkpoints/step_2000.pt

The zero-byte training lock file .step_2000.pt.lock is intentionally not included. The actual checkpoint, model source, processor assets, dataset statistics, and offline Python/CUDA dependencies are included.

The official InferenceNode.load_model() and InferenceNode.predict() hooks are implemented. predict() returns a finite float32 NumPy array with shape (32, 25).

Load

Either archive can be loaded; the gzip archive is smaller:

docker load -i QQ.tar.gz

Run

docker run --rm --gpus all --network host --ipc host -it \
  g05-newdata-step2000:iros2026-fa2-fmonly-pathfix /bin/bash

Inside the container:

bash scripts/run_infer.sh

The deployment action path is FM-only by default (G05_ENABLE_AR_ACTION=0). It replans a 32-step action chunk at approximately 2 Hz while the independent publisher sends steps at 30 Hz; a newly inferred trajectory replaces the remaining old trajectory. For diagnosis only, set G05_ENABLE_AR_ACTION=1 to restore AR+FM inference; this substantially reduces the replanning rate.

eval/infer_setting.py is the single runtime source for ckpt_path and config_path. The corresponding optional environment overrides are G05_CHECKPOINT and G05_CONFIG_PATH; both resolved paths are validated and printed before model loading.

The model runs fully offline (HF_HUB_OFFLINE=1 and TRANSFORMERS_OFFLINE=1). FlashAttention 2.8.3.post1 is installed and used by the vision encoder. GitPython's optional repository probe is disabled with GIT_PYTHON_REFRESH=quiet; model inference does not use git.

Validation performed

• Official base archive SHA-256 verified before import.
• Final Docker archive manifest parsed successfully: linux/amd64.
• NVIDIA entrypoint and official ROS environment preserved.
• Embedded checkpoint SHA-256 matches the source checkpoint.
• Container-root imports passed for Torch, NumPy, OpenCV, Transformers, Hydra, G05, and the challenge adapter.
• Real checkpoint load and synthetic (32, 25) inference passed on an A100 from the assembled image root filesystem with host GPU devices and driver libraries injected as a container runtime would provide them.

Before an official submission, run the image once on the competition-equivalent RTX 4090 48 GB host and follow any final upload/naming instructions announced by the organizers.