Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LiveNet: Robust, Minimally Invasive Multi-Robot Control for Safe and Live Navigation in Constrained Environments

Srikar Gouru, Siddharth Lakkoju, Rohan Chandra

TL;DR

LiveNet addresses safe and live navigation for multiple robots operating in constrained environments by embedding differentiable control barrier functions (dCBFs) into a fully decentralized neural controller. It achieves minimally invasive, deadlock-free motion without inter-agent communication, with per-cycle compute speed between $10$-$20\times$ faster than MPC-based methods and about $20\times$ faster than MACBF, and demonstrates provable liveness alongside safety. The approach integrates a barrier-function-based QP (OptNet) layer within BarrierNet-like architecture, learning penalty relaxations for obstacle avoidance and liveness to yield smooth, human-like trajectories that adapt to varied scenario configurations. Open-source code is provided, and results show LiveNet outperforms baselines in doorway and intersection SMGs, offering practical potential for real-world deployment and future work on scalability and unsupervised learning to reduce data-generation costs.

Abstract

Robots in densely populated real-world environments frequently encounter constrained and cluttered situations such as passing through narrow doorways, hallways, and corridor intersections, where conflicts over limited space result in collisions or deadlocks among the robots. Current decentralized state-of-the-art optimization- and neural network-based approaches (i) are predominantly designed for general open spaces, and (ii) are overly conservative, either guaranteeing safety, or liveness, but not both. While some solutions rely on centralized conflict resolution, their highly invasive trajectories make them impractical for real-world deployment. This paper introduces LiveNet, a fully decentralized and robust neural network controller that enables human-like yielding and passing, resulting in agile, non-conservative, deadlock-free, and safe, navigation in congested, conflict-prone spaces. LiveNet is minimally invasive, without requiring inter-agent communication or cooperative behavior. The key insight behind LiveNet is a unified CBF formulation for simultaneous safety and liveness, which we integrate within a neural network for robustness. We evaluated LiveNet in simulation and found that general multi-robot optimization- and learning-based navigation methods fail to even reach the goal, and while methods designed specially for such environments do succeed, they are 10-20 times slower, 4-5 times more invasive, and much less robust to variations in the scenario configuration such as changes in the start states and goal states, among others. We open-source the LiveNet code at https://github.com/srikarg89/LiveNet{https://github.com/srikarg89/LiveNet.

LiveNet: Robust, Minimally Invasive Multi-Robot Control for Safe and Live Navigation in Constrained Environments

TL;DR

LiveNet addresses safe and live navigation for multiple robots operating in constrained environments by embedding differentiable control barrier functions (dCBFs) into a fully decentralized neural controller. It achieves minimally invasive, deadlock-free motion without inter-agent communication, with per-cycle compute speed between - faster than MPC-based methods and about faster than MACBF, and demonstrates provable liveness alongside safety. The approach integrates a barrier-function-based QP (OptNet) layer within BarrierNet-like architecture, learning penalty relaxations for obstacle avoidance and liveness to yield smooth, human-like trajectories that adapt to varied scenario configurations. Open-source code is provided, and results show LiveNet outperforms baselines in doorway and intersection SMGs, offering practical potential for real-world deployment and future work on scalability and unsupervised learning to reduce data-generation costs.

Abstract

Robots in densely populated real-world environments frequently encounter constrained and cluttered situations such as passing through narrow doorways, hallways, and corridor intersections, where conflicts over limited space result in collisions or deadlocks among the robots. Current decentralized state-of-the-art optimization- and neural network-based approaches (i) are predominantly designed for general open spaces, and (ii) are overly conservative, either guaranteeing safety, or liveness, but not both. While some solutions rely on centralized conflict resolution, their highly invasive trajectories make them impractical for real-world deployment. This paper introduces LiveNet, a fully decentralized and robust neural network controller that enables human-like yielding and passing, resulting in agile, non-conservative, deadlock-free, and safe, navigation in congested, conflict-prone spaces. LiveNet is minimally invasive, without requiring inter-agent communication or cooperative behavior. The key insight behind LiveNet is a unified CBF formulation for simultaneous safety and liveness, which we integrate within a neural network for robustness. We evaluated LiveNet in simulation and found that general multi-robot optimization- and learning-based navigation methods fail to even reach the goal, and while methods designed specially for such environments do succeed, they are 10-20 times slower, 4-5 times more invasive, and much less robust to variations in the scenario configuration such as changes in the start states and goal states, among others. We open-source the LiveNet code at https://github.com/srikarg89/LiveNet{https://github.com/srikarg89/LiveNet.

Paper Structure

This paper contains 13 sections, 1 theorem, 13 equations, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Simultaneous Safety and Deadlock Resolution Methods
  4. Other Multi-Robot Navigation Approaches
  5. Problem Formulation
  6. LiveNet: Technical Approach
  7. Multi-Agent Collision Avoidance dCBFs
  8. Minimally Invasive Deadlock Prevention dCBFs
  9. Experiments and Results
  10. Experiment Setup
  11. Training:
  12. Results
  13. Conclusion

Key Result

theorem 1

Assuming $p^l_{-1}(z), p^l_1(z)$ are differentiable functions with respect to $z$, then the LiveNet constraints in Equation eqn: faster_agent_live_CBF guarantee the liveness of the system defined by eq: multi-agent-state-space-system-equation.

Figures (6)

  • Figure 1: LiveNet enables minimally invasive, robust, safe and deadlock-free navigation in constrained environments compared to existing methods.
  • Figure 2: Example SMG and Non-SMG scenarios with agent 1's desired trajectory in red and agent 2's desired trajectory in blue. Their starting and goal locations are indicated by $t = 1$ and $t = 4$, respectively, with $t$ being used to show the agents' time-parameterized desired trajectories. Collisions are shown in purple.
  • Figure 3: LiveNet Architecture Overview The ego state and observation inputs get fed into a feedforward network with three individual outputs: the reference control, the obstacle penalties, and the minimal invasiveness penalties. These three outputs get fed into a differentiable QP layer which solves a standard QP problem with inequality constraints (Equations \ref{['eqn:penalty_lie_ineq']} and \ref{['eq: liveness_penalty_lie_ineq']}) to enforce the CBFs. During backpropagation, the optimal reference value, as well as optimal penalty values for the CBF constraints, are learned.
  • Figure 4: Resulting trajectories in Doorway (Figures \ref{['fig: mpc_cbf_doorway_scenario']}-\ref{['fig: livenet_doorway_scenario']}) and Intersection (Figures \ref{['fig: mpc_cbf_intersection_scenario']}-\ref{['fig: livenet_intersection_scenario']})
  • Figure 5: LiveNet's obstacle dCBF, liveness dCBF, and deviation from desired path in Doorway scenario.
  • ...and 1 more figures

Theorems & Definitions (2)

  • theorem 1
  • proof