RO-50-4eph-trial1-repo

RO-Inference

Integrated robot-origami inference stack:

1. OOI cropper
2. Checkpoint planner
3. OpenPI VLA
4. Public Zenoh server aligned with the participant kit

Entry point:

• E:/Robot-Origami-Challenge/RO-Inference/serve_origami_policy.py

Main runtime config:

• E:/Robot-Origami-Challenge/RO-Inference/configs/origami_inference.yaml

Model files to place

OOI checkpoint:

• E:/Robot-Origami-Challenge/RO-Inference/models/ooi/best.pt

Checkpoint planner checkpoint:

• E:/Robot-Origami-Challenge/RO-Inference/models/checkpoint_planner/step_0015000.pt

DINOv3 model directory:

• E:/Robot-Origami-Challenge/RO-Inference/models/dino/dinov3-vits16plus-pretrain-lvd1689m

OpenPI checkpoint bundle and tokenizer:

• E:/Robot-Origami-Challenge/RO-Inference/OpenPI_Module/model_bundles/pi05_origami_checkpoint_spline_vla
• E:/Robot-Origami-Challenge/RO-Inference/OpenPI_Module/assets/paligemma_tokenizer.model

Runtime modes

• input_mode=compressed_224 Use the participant-kit public image size. OOI and planner internally upscale to 480, while the VLA keeps the original 224 camera images and uses a 224 resize of the OOI crop.

• input_mode=native_480 Local-only mode for native 480x480 camera inputs. The public competition contract still uses 224x224, so this is mainly for controlled internal runs.

• pipeline_mode=end_to_end Single process with PyTorch OOI/planner plus JAX OpenPI on one GPU.

• pipeline_mode=separate_servers Single public Zenoh server, but two internal worker processes:

  • PyTorch worker for OOI + checkpoint planner
  • JAX worker for OpenPI VLA

Example

cd E:\Robot-Origami-Challenge\RO-Inference
python .\serve_origami_policy.py `
  --config .\configs\origami_inference.yaml `
  --endpoint tcp/127.0.0.1:7447 `
  --session-id test-session `
  --input-mode compressed_224 `
  --pipeline-mode end_to_end `
  --speed 1.0 `
  --num-sampled-actions 20 `
  --policy-sample-steps 10 `
  --planner-transition-threshold 0.75

Submission bundle checks

Verify that the bundle is fully self-contained and offline-safe:

cd E:\Robot-Origami-Challenge\RO-Inference
python .\scripts\verify_submission_bundle.py --config .\configs\origami_inference.yaml

This checks:

• OOI checkpoint
• checkpoint-planner checkpoint
• local DINOv3 dinov3-vits16plus-pretrain-lvd1689m
• OpenPI checkpoint bundle
• paligemma_tokenizer.model
• norm_stats.json
• origami_tactile_norm_stats.json
• that all runtime paths stay inside RO-Inference
• that DINO is configured for local offline loading

Submission container

Docker assets are provided in:

• E:/Robot-Origami-Challenge/RO-Inference/docker/submission-zenoh-bundled.Dockerfile
• E:/Robot-Origami-Challenge/RO-Inference/docker/submission_zenoh_entrypoint.sh
• E:/Robot-Origami-Challenge/RO-Inference/requirements.submission.txt

Build from the RO-Inference directory as the Docker context:

cd E:\Robot-Origami-Challenge\RO-Inference
docker build -f .\docker\submission-zenoh-bundled.Dockerfile -t ro-inference-submission .

The container entrypoint requires:

• ORIGAMI_ZENOH_ENDPOINT
• ORIGAMI_SESSION_ID

Optional startup knobs:

• ORIGAMI_SPEED
• ORIGAMI_NUM_SAMPLED_ACTIONS
• ORIGAMI_PLANNER_TRANSITION_THRESHOLD
• ORIGAMI_POLICY_SAMPLE_STEPS
• ORIGAMI_INPUT_MODE
• ORIGAMI_PIPELINE_MODE

Default submission mode is compressed_224. native_480 remains available as a local debug mode.