Pareto efficiency in synthesizing shared autonomy policies with temporal logic constraints

Abstract

Abstract-In systems in which control authority is shared by an autonomous controller and a human operator, it is important to find solutions that achieve a desirable system performance with a reasonable workload for the human operator. We formulate a shared autonomy system capable of capturing the interaction and switching control between an autonomous controller and a human operator, as well as the evolution of the operator's cognitive state during control execution. To trade-off human's effort and the performance level, e.g., measured by the probability of satisfying the underlying temporal logic specification, a two-stage policy synthesis algorithm is proposed for generating Pareto efficient coordination and control policies with respect to user specified weights. We integrate the Tchebychev scalarization method for multi-objective optimization methods to obtain a better coverage of the set of Pareto efficient solutions than linear scalarization methods.

Quantifying Hypothesis Space Misspecification in Learning From Human–Robot Demonstrations and Physical Corrections

“…4c relates the induced feature values Φ D to θ as a function of the E. When E = 1, the induced features will tend to have low cost with respect to θ; when E = 0, the induced features do not depend on θ, and we model them as Gaussian noise centered around the feature values of the robot's currently planned trajectory x R . 5 Note thatβ is non-negative, since u * H is the minimal-normũ H that satisfies the constraint, so the difference in the denominator of (24) is positive. with the constant in the E = 0 case corresponding to the normalization term of the normal distribution.…”

“…In order to utilize human input, system designers typically equip robots with a representation of possible objectives that the human could care about. These representations can range from quadratic cost models [4] to complex temporal logic specifications [5] to neural networks [6]. However, anticipating all motivations for human input and specifying a complete model is challenging.…”

Human-in-the-Loop Mixed-Initiative Control Under Temporal Tasks

“…Now we show the computational efficiency of Alg. 1 compared with two straight-forward solutions: (M1) choose the optimal β among a set of guessed values of β, denoted by S β ; (M2) solve (16) directly by enumerating all runs in R q1q H . The first method's accuracy relies on S β being large, which however results in high computational cost.…”

Trust-Based Multi-Robot Symbolic Motion Planning with a Human-in-the-Loop

“…Since the explored system state needs to be stored in memory, this makes the problem intractable for systems with realistic size. Hence, there has been limited work on symbolic motion planning for fully autonomous multi-agent systems [2,37,60,64] and very few recent results on HRI systems based on formal verification [20,24,47]. The papers [37,64] investigate the task decomposition problem to address the scalability in multi-agent planning.…”

“…Also note that here we focus on the high-level switching strategies; the low-level continuous robot motion execution under either manual or autonomous motion planning is still automatic. This differs from the literature on automaton-based switching/shared control synthesis for human-robot teams [24,47]; (4) Deadlock-and livelock-free algorithms are proposed to guarantee reachability of all the goals with a human-in-the-loop. We provide a formal proof for the correctness of the overall algorithm and motivate its need through a discussion of inter-robot collision scenarios that cannot be resolved by purely using the methods summarized in (2); (5) We perform non-trivial multi-robot simulations with direct human inputs (e.g., through gamepad and mouse with GUI designs) to demonstrate the real-time trust computation and the implementation of the proposed trust-based symbolic robot motion planning methods.…”