Robots stop asking the big model every step: a local CALL head pages Astra only for recovery, lifting 200 LIBERO episodes to 97.5%

A robot manipulation policy can now decide for itself when it needs help from an external reasoning model, instead of querying one at every step. The earlier hybrid scheme had Astra continuously review and correct policy proposals, incurring overhead even when the underlying proposal was correct. Across 40 tasks and 200 episodes in the LIBERO-10, Object, Goal, and Spatial suites, the hybrid system succeeded in 195/200 episodes versus 191/200 for pure π0.5.

The decision is made locally, reusing representations the policy already computes, with no extra model invocation. The CALL head concatenates a semantic feature, an action feature, and a 44-dimensional vector of robot state plus action-chunk statistics, then projects them through LayerNorm, a 256-unit linear layer, GELU, LayerNorm, and a scalar output, for 479,937 trainable parameters. Training data combines failure states from LIBERO-Recovery with normally progressing states, split by task into training, early stopping, calibration, and testing sets, totaling 5,055 competence-state examples over 37 held-out task folds.

Control passes to Astra only when help is confirmed, and returns to π0.5 once the state stabilizes. CALL is evaluated at each policy query, consuming five actions from the policy chunk between queries during normal execution, and a five-query score window requires at least four scores above the calibrated threshold. Of 200 episodes, 61 consulted Astra and 13 executed Astra actions; in the Object, Goal, and Spatial suites Astra executed 1,076 of 19,592 environment steps, a 5.49% share.

The system is evaluated on 40 tasks and 200 episodes across the LIBERO-10, Object, Goal, and Spatial suites, with a frozen OpenPI pi05_libero checkpoint as the base policy and Astra recovering from external and wrist RGB images, the task instruction, robot state, the policy proposal, and execution history. It suits readers tracking inference cost and recovery mechanisms in embodied policies.

The Object suite shows a 38% consultation rate while pure π0.5 continuation succeeds in 95.79% of trials at first-CALL states, indicating frequent unnecessary triggers; the authors attribute this to the suite's lower difficulty, homogeneous scenes, and an overrepresentation of failure states in its training data, and propose refining the CALL module and rebalancing its training data.