Technical Summary: Principled Authority Switching for Shared Autonomy in Human-Robot Teams

Problem Statement

Shared autonomy in cyber-physical systems (CPS) requires mechanisms to dynamically allocate control between humans and autonomous agents. Existing approaches often rely on blending control inputs or heuristic switching rules. These methods suffer from a lack of theoretical guarantees, dependence on accurate intent prediction, and the imposition of continuous cognitive burdens on human operators (requiring constant input). Furthermore, they frequently fail to account for the human's actual likelihood of intervention (override propensity), leading to suboptimal coordination in safety-critical domains such as autonomous driving and teleoperation.

The core challenge addressed is how to formulate authority switching as a cooperative, team-optimal decision problem that explicitly models the human's override capability and the costs associated with switching control, rather than relying on ad hoc rules.

Methodology: The Flip-Team Framework

The authors propose Flip-Team, a cooperative framework that models authority switching as an identical-interest dynamic game. The system operates under a "human-on-the-loop" paradigm, where the autonomous system operates independently, and the human retains supervisory override authority.

1. System Dynamics and State

The system is modeled as a discrete-time dynamical system where the control authority at time $k$ is denoted by a FlipDyn state $\alpha_k \in \{H, A\}$ (Human or Autonomous).

• Human Control: $x_{k+1} = F^H_k(x_k, u_k)$
• Autonomous Control: $x_{k+1} = F^A_k(x_k, w_k)$
• Takeover Actions: Agents take actions $\pi^j_k \in \{0, 1\}$ (idle or request takeover).
• Transition Logic: Transitions are mutually exclusive. Crucially, when the autonomous agent is in control, a human override attempt succeeds with a probability $p_k$ (the human's override propensity) if the autonomous agent does not request a handoff. If the autonomous agent requests a handoff, the transition is deterministic.

2. Cost Structure

The objective is to minimize a total cost function $J$ over a finite horizon $L$:
$$J = g_{L+1}(x_{L+1}, \alpha_{L+1}) + \sum_{t=1}^{L} \left( g_t(x_t, \alpha_t) + \pi^H_t h_t(x_t) + \pi^A_t a_t(x_t) \right)$$

• State Cost ($g_t$): Represents performance metrics (e.g., tracking error, energy) which differ between human and autonomous control ($g^H_t \neq g^A_t$).
• Switching Costs ($h_t, a_t$): Represent the cognitive load, attention switching, or transition risk for the human and autonomous agent, respectively.

3. Optimal Switching Policy (Linear-Quadratic Case)

For linear dynamics with quadratic costs (LQ systems), the authors derive closed-form solutions using dynamic programming.

• Value Functions: The cost-to-go is represented by quadratic value functions $V^H_k(x) = x^\top P^H_k x$ and $V^A_k(x) = x^\top P^A_k x$.
• Recursions: The matrices $P^H_k$ and $P^A_k$ are computed via backward recursion.
• Theorem 1 (Switching Conditions): The optimal policy is determined by comparing the relative cost advantage of human vs. autonomous control (quantified by $\tilde{P}_{k+1}$) against the switching costs and the override probability $p_k$.
  • When Human is in Control: Retains control if the cost advantage of human control plus switching costs is favorable; otherwise, a coordinated handoff to autonomy occurs.
  • When Autonomous is in Control: Three regimes emerge based on $p_k$:
    1. Retain: If the cost advantage of human control is insufficient relative to $p_k$.
    2. Override: If the cost advantage is significant and $p_k$ is high enough, the human unilaterally takes over.
    3. Coordinated Handoff: If the cost advantage is large, both agents agree to transfer control.

4. Override Threshold and Estimation

• Theorem 2 (Critical Threshold): The paper derives a critical threshold $p^*_k(x)$ that determines when human intervention is cost-effective. For isotropic switching costs, this simplifies to a state-independent ratio: $p^* = h / (h + a)$.
  • If the human's actual propensity $p_k < p^*$, intervention is likely not cost-effective.
  • If $p_k > p^*$, intervention is justified.
• Online Estimation: Since $p_k$ is unknown in practice, the authors propose an online estimation method. This involves tracking the frequency of human overrides during specific conditions (autonomous control, no handoff request) across episodes or using exponential smoothing within an episode to adaptively estimate $\hat{p}_k$.

Key Contributions

1. Principled Switching Framework: Formulation of authority switching as a cooperative game with asymmetric costs and human override capability, yielding optimal policies without heuristic tuning.
2. Override Threshold Derivation: Derivation of a critical threshold $p^*_k$ that dictates when human takeover is cost-effective, providing interpretable design guidelines.
3. Closed-Form Solutions: For linear-quadratic systems, the derivation of closed-form switching conditions and value function recursions that enable efficient computation independent of the continuous state dimension.
4. Adaptive Estimation: A proposed method to estimate override propensity online from observed interventions, enabling adaptive switching without prior calibration.

Evaluation and Results

The framework was evaluated on a 1D quadrotor altitude regulation task (double integrator dynamics) with a horizon of 30 steps.

• Baselines: The Flip-Team policy was compared against:
  1. Always-autonomous.
  2. Always-human.
  3. A performance-threshold heuristic (switching based solely on tracking error).
• Metrics: Total cost, number of switches, and percentage of time under human control.
• Results:
  • Flip-Team achieved the lowest total cost (40.5), outperforming the always-human baseline (43.2) by 6% and the always-autonomous baseline (45.6) by 11%.
  • The performance-threshold baseline incurred the highest cost (55.3), demonstrating that reactive switching based on error alone, without considering propensity or switching costs, degrades performance.
  • Flip-Team achieved these results with only 1.8 authority switches on average, with the human in control for 56% of the time.
  • The optimal policy successfully anticipated phases where human intervention was both likely (high $p_k$) and beneficial (cost advantage), avoiding unnecessary switches during low-engagement phases.

Significance and Claims

The paper claims that Flip-Team provides a principled mechanism for shared autonomy that balances human adaptability with autonomous efficiency. By explicitly modeling the human's override propensity and the costs of switching, the framework avoids the pitfalls of heuristic rules that either overburden the human or fail to intervene when necessary.

The authors emphasize that the derived critical threshold offers actionable design guidelines:

• Designers can tune switching costs ($H_k, A_k$) to shape the engagement landscape.
• Practitioners can modulate the human's propensity (via alerts or trust calibration) to ensure it remains above the critical threshold when intervention is needed.

The work is modest in its current scope, noting that the evaluation relies on a simulated human model with a hand-crafted propensity profile and a 1D task. The authors state that future work will validate the framework with human-in-the-loop experiments, extend to state-dependent propensity estimation, and apply the framework to higher-dimensional and nonlinear systems.