PASA-KV

PASA-KV · Evaluation

Results and diagnostics

Policy success is the robot's task success rate. FastWAM training used 40 tasks across LIBERO Spatial, Object, Goal, and LIBERO-10. The primary closed loop evaluation uses LIBERO-10: 10 tasks with 50 trials each

Simulation results ↓Route selection ↓Piper robot results ↓

Primary comparison

Policy success and recorded GPU hours needed to construct each patch, as reported in the paper. Each model uses its own cost scale

Policy success rates and recorded patch construction cost across FastWAM, LingBot-VA, and Cosmos Policy
Paper Table 1. Lower policy success indicates a stronger attack. The cost panels report patch construction time

← Project page simulation overview

Route selection

Matched FastWAM controls use the same attack loss, patch geometry, target count, and training budget. This comparison uses 20 trials per LIBERO-10 task

Matched FastWAM route selection controls with Tail-AV at 1.5 percent policy success
Selection control. Tail-AV selects routes from action readouts without using rollout success labels

Piper robot outcomes

Each reported condition has 10 physical trials. Baseline comparisons cover bowl stacking only and use different patch geometries

Piper task success counts and Wilson intervals across three tasks and bowl stacking baselines
Real robot results. Wilson 95% intervals show uncertainty from the reported success counts. These intervals were calculated for this project page and are not part of the paper tables

← Project page videos and real robot overview