AdaRoute Release Artifacts

AdaRoute combines a frozen manipulation policy with a learned planning trigger, finite motion-planner endpoints, and an action-conditioned world model. Code and exact reproduction instructions are at Loule0-0/AdaRoute.

Contents

• start_trigger/release_v1/: current best all-task start-trigger adapter, frozen configuration, metrics, training history, and saved validation/test predictions.
• start_trigger/ood_shape_wall_eval/: strict leave-wall-out checkpoint retained for the obstacle-shape OOD ablation.
• acwm/task8_lora/: narrow task-8 WoVR LoRA smoke checkpoint.
• SHA256SUMS: integrity manifest for every binary and experiment output above.

Start-trigger inputs and target

The release adapter reads the instruction, frozen pi0.5 dual-view 8x8 spatial prefix features, 22-D state, and the exact first eight raw 7-D actions from the current pi0.5 query. A deterministic cumulative transform expands each step to 15 intent features: raw action (7), cumulative translation (3), cumulative rotation (3), translation-step norm (1), and rotation-step norm (1), followed by normalization. Planner bridges must be stored in the same raw 7-D OSC action convention before this transform. It predicts whether a physical blocking event will occur within 48 control steps, with an eight-step minimum useful notice window.

Simulator contact state and obstacle masks are used only to build offline labels and evaluation targets. They are not online inputs.

Results and limitation

At the global validation-selected threshold, held-out test valid-trigger recall is 39.3%, worst-task recall is 27.3%, and safe-trajectory false-trigger rate is 5.4%. Task-conditional calibration gives 46.4%, 33.3%, and 0%, respectively. Sample ROC-AUC is 0.923.

This is the current best loadable all-task checkpoint, but it is not deployment-ready and should not be interpreted as a high-success real-robot trigger. The task-8 WoVR LoRA is trained from only four trajectories and exists to reproduce the early closed-loop smoke chain, not to claim AC-WM generalization.

Data and reproduction

The exact raw rollouts, immutable split arrays, GT masks, and pi0.5 feature shards are in Jaqen0-0/AdaRoute under start_hazard_v4diverse/.

git clone https://github.com/Loule0-0/AdaRoute.git
cd AdaRoute
bash scripts/download_release_artifacts.sh checkpoints/adaroute_release

See docs/reproducibility.md for pinned source revisions, container digest, package snapshots, data checksums, and the exact training command.