When should a robot ask for help?

Robotics: when should a robot ask for help? (simulated people)

A robot must fetch “the red cup on the left”. Before acting it can ask two clarification questions; the answers of a simulated population (quantum-like order effects, illustrative parameters) arrive in batches. After each batch the robot refits its HumanModelEnsemble, predicts how likely a person is to confirm its hypothesis, and uses ask_or_act() to decide whether asking is worth its cost. While the human models still disagree the robot asks whatever the costs; once they agree, the decision is driven by the cost of a wrong action: with an expensive error (3) it keeps asking, with a cheap error (1.5) it acts.

This is a decision layer only: no robot middleware is needed. In a ROS 2 system, the same calls would sit in a service or behaviour-tree node (not included in the package).

import numpy as np
from quantum_mind.applications.robotics import domain_models, HumanModelEnsemble, ask_or_act
from quantum_mind.families.order_effects import QuantumOrderModel4D, BayesOrderModel, AnchoringOrderModel

Answers arrive in batches

After each batch the ensemble is refitted and the robot decides, for two error costs, whether to ask or to act.

rng = np.random.default_rng(3)
people = domain_models('object_clarification')['QL']                # simulated population
counts = {'AB': np.zeros(4, int), 'BA': np.zeros(4, int)}
print('batch  people  P(confirm)  disagreement(bits)  error cost 3.0   error cost 1.5')
for batch in range(1, 7):
    new = people.sample(None, 25, rng)
    counts = {c: counts[c] + new[c] for c in counts}
    ens = HumanModelEnsemble([QuantumOrderModel4D, BayesOrderModel, AnchoringOrderModel]).update(counts, restarts=4, rng=rng)
    p, unc = ens.predict('AB')
    p_confirm = p[0] + p[1]                                            # 'yes' to the first question
    hi, lo = (ask_or_act(p_confirm, unc, ask_cost=1.0, error_cost=c, max_disagreement_bits=0.01)['action'] for c in (3.0, 1.5))
    print('%5d  %6d  %10.3f  %18.4f  %-15s %s' % (batch, 50 * batch, p_confirm, unc['model_disagreement_bits'], hi, lo))
print('ensemble weights:', [(s['model'], round(s['weight'], 2)) for s in ens.summary()])

batch  people  P(confirm)  disagreement(bits)  error cost 3.0   error cost 1.5
    1      50       0.443              0.0225  ask             ask
    2     100       0.437              0.0016  ask             act
    3     150       0.501              0.0024  ask             act
    4     200       0.524              0.0015  ask             act
    5     250       0.516              0.0008  ask             act
    6     300       0.517              0.0010  ask             act
ensemble weights: [('BayesOrderModel', 0.77), ('QuantumOrderModel4D', 0.23), ('AnchoringOrderModel', 0.0)]