Skip to main content

CA Current Framework#

Current implementation of Collision Avoidiance for the multi-agent UAVs is using tyhe velocity obstacles implementation. This relies on the relative velocity vector and Protected Zone to prevent collision.

Refer to paper :

Y. Kuwata, M. T. Wolf, D. Zarzhitsky and T. L. Huntsberger, "Safe maritime navigation with COLREGS using velocity obstacles," 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2011, pp. 4728-4734, doi: 10.1109/IROS.2011.6094677. https://ieeexplore.ieee.org/document/6094677

Problems Identified from Current CA#

  • Collisions occur, especially when congested​

  • Unnecessary avoidance, disrupting mission​

  • Not following rules properly all the time​

  • Unfairness, some drones do all the work to avoid​

  • Deadlock in tight spaces​

  • Shallow crossing encounter, the two drones struggle to pass each other and could go into deadlock​

  • Simplify the parameters (update rate, latency, detection range, etc) – too many and they are actually related and not easy to determine what value to assign​

  • Priority issue – when a drone is given highest priority, are we sure others would be able to avoid it while it ignores everyone else​

  • Handling static obstacles​

Conjectures of New CA#

  1. Adjustment of trajectory in a mutually agreed opposite direction (pass on the right side rule) will converge stably in a non-oscillatory manner when the maneuver is half or less than what is required to avoid => deadlock won’t occur because the state of deadlock is an unstable equilibrium​

  2. A resolution is always possible because the drones are assumed to start on a collision-free path and they would not change their trajectories unless it is collision-free (eventually – because they only change by half or less than what is required)​

  3. The model to approximate the trajectory of the drone is representative of the actual trajectory => violation of PZ will only be because of failing to use a good model of the drone and associated buffer for uncertainties​

  4. Gradient descent – there is a no local minimum in the cost function ​

Collision Test Cases#

Basic :#

Passing Static UAVCrossingHead OnOvertakingShallow Crossing
Passing Static UAVCrossingHead OnOvertakingShallow Corssing

Advance :#

AntipodalLawn MowingOrbitNarrow Spaces
AntipodalLawn MowingOrbitNarrow Spaces

Considerations for implementation#

  • When no solution is found, we stop the drone. Note that stopping would not work if other drone is higher priority and does not try to avoid it​

  • This algo may require a low level, more reactive algo as a last line of defence to avoid others. To think about this later and to define its role and for what cases it is needed. It could be:​

    • The current CA algo itself​
    • A simple repulsion rule​

Optimization of implementation#

  • Collision Check : May not need to propagate with small time steps and there may be a smarter way to do this more efficiently.​
  • Fairness in avoidance : We can exchange costs and moderate manAmt as a function of my cost vs your cost​.
Last updated on