VLAarchtests/artifacts/reports/task_routed_proxy_v1/summary.md

Task-Routed Proxy Controller

• routing rule: foliage -> iter6, bag -> iter8, cloth -> iter8
• fixed benchmark slices: 100 episodes per task from the standard sprint spec
• mean success: 0.4867

Per-Task Success

• foliage: 0.46
• bag: 0.41
• cloth: 0.59

Why This Matters

• This is the strongest current controller for the three custom proxy tasks.
• The routing rule is fair: it depends only on explicit task metadata, not on per-episode oracle information.
• Compared with the best single checkpoint (iter7/iter8 at 0.4667), the routed controller keeps the cloth gain while recovering the stronger foliage checkpoint.

Source Runs

• foliage source: /workspace/VLAarchtests/artifacts/reports/selector_finetune_v7_iter6/default/reveal_benchmark.json
• bag source: /workspace/VLAarchtests/artifacts/reports/selector_finetune_v7_iter8/bag_fixed_default/reveal_benchmark.json
• cloth source: /workspace/VLAarchtests/artifacts/reports/selector_finetune_v7_iter8/cloth_fixed_default/reveal_benchmark.json

Reproduction

• benchmark wrapper: /workspace/VLAarchtests/code/reveal_vla_bimanual/scripts/run_task_routed_proxy_eval.sh
• routing is now supported directly in run_reveal_benchmark.py through --task-routed-model