Model Predictive Path Integral Controller

:rotating_light: :rotating_light: While this project repository is absolutely critical and led to MPPI's inclusion into the Nav2 framework by the amazing work and efforts by @artofnothingness, the version in Nav2's repository fixes a number of bugs major and minor and will continue to receive community improvements and be maintained alongside the rest of the navigation stack into the future. Please see that version moving forward! This repository continues to be an important historical project to reference decisions made and Git histories of changes during development. :rotating_light: :rotating_light:

Overview

This is a predictive controller (local trajectory planner) that implements the Model Predictive Path Integral (MPPI) algorithm to track a path with adaptive collision avoidance. It contains plugin-based critic functions to impact the behavior of the algorithm. It was created by Aleksei Budyakov and adapted & developed for Nav2 by Steve Macenski.

This plugin implements the nav2_core::Controller interface allowing it to be used across the navigation stack as a local trajectory planner in the controller server's action server (controller_server).

This controller is measured to run at 50+ Hz on a modest Intel processor (4th gen i5). See its Configuration Guide Page for additional parameter descriptions.

It works currently with Differential, Omnidirectional, and Ackermann robots.

MPPI Description

The MPPI algorithm is an MPC variant that finds a control velocity for the robot using an iterative approach. Using the previous time step's best control solution and the robot's current state, a set of randomly sampled perturbations from a Gaussian distribution are applied. These noised controls are forward simulated to generate a set of trajectories within the robot's motion model.

Next, these trajectories are scored using a set of plugin-based critic functions to find the best trajectory in the batch. The output scores are used to set the best control with a soft max function.

This process is then repeated a number of times and returns a converged solution. This solution is then used as the basis of the next time step's initial control.

Features

• Predictive MPC trajectory planner
• Utilizes plugin-based critics which can be swapped out, tuned, or replaced easily by the user
• Highly optimized CPU-only performance using vectorization and tensor operations
• Supports a number of common motion models, including Ackermann, Differential-Drive, and Omni-directional
• Includes fallback mechanisms to handle soft-failures before escalating to recovery behaviors
• High-quality code implementation with Doxygen, high unit test coverage, documentation, and parameter guide
• Easily extensible to support modern research variants of MPPI

Configuration

Controller

Trajectory Visualizer

Path Handler

Ackermann Motion Model

Constraint Critic

Goal Angle Critic

Goal Critic

Obstacles Critic

Path Align Critic

Path Angle Critic

Path Follow Critic

Prefer Forward Critic

Twirling Critic

XML configuration example

controller_server:
  ros__parameters:
    controller_frequency: 30.0
    FollowPath:
      plugin: "mppi::MPPIController"
      time_steps: 56
      model_dt: 0.05
      batch_size: 2000
      vx_std: 0.2
      vy_std: 0.2
      wz_std: 0.4
      vx_max: 0.5
      vx_min: -0.35
      vy_max: 0.5
      wz_max: 1.9
      iteration_count: 1
      prune_distance: 1.7
      transform_tolerance: 0.1
      temperature: 0.3
      gamma: 0.015
      motion_model: "DiffDrive"
      visualize: false
      TrajectoryVisualizer:
        trajectory_step: 5
        time_step: 3
      AckermannConstrains:
        min_turning_r: 0.2
      critics: ["ConstraintCritic", "ObstaclesCritic", "GoalCritic", "GoalAngleCritic", "PathAlignCritic", "PathFollowCritic", "PathAngleCritic", "PreferForwardCritic"]
      ConstraintCritic:
        enabled: true
        cost_power: 1
        cost_weight: 4.0
      GoalCritic:
        enabled: true
        cost_power: 1
        cost_weight: 5.0
        threshold_to_consider: 1.0
      GoalAngleCritic:
        enabled: true
        cost_power: 1
        cost_weight: 3.0
        threshold_to_consider: 0.4
      PreferForwardCritic:
        enabled: true
        cost_power: 1
        cost_weight: 5.0
        threshold_to_consider: 0.4
      ObstaclesCritic:
        enabled: true
        cost_power: 1
        repulsion_weight: 1.5
        critical_weight: 20.0
        consider_footprint: false
        collision_cost: 10000.0
        collision_margin_distance: 0.1
        near_goal_distance: 0.5
      PathAlignCritic:
        enabled: true
        cost_power: 1
        cost_weight: 14.0
        max_path_occupancy_ratio: 0.05
        trajectory_point_step: 3
        threshold_to_consider: 0.40
        offset_from_furthest: 20
      PathFollowCritic:
        enabled: true
        cost_power: 1
        cost_weight: 5.0
        offset_from_furthest: 5
        threshold_to_consider: 0.6
      PathAngleCritic:
        enabled: true
        cost_power: 1
        cost_weight: 2.0
        offset_from_furthest: 4
        threshold_to_consider: 0.40
        max_angle_to_furthest: 1.0
      # TwirlingCritic:
      #   enabled: true
      #   twirling_cost_power: 1
      #   twirling_cost_weight: 10.0

Topics

Notes to Users

General Words of Wisdom

The model_dt parameter generally should be set to the duration of your control frequency. So if your control frequency is 20hz, this should be 0.05. However, you may also set it lower but not larger.

Visualization of the trajectories using visualize uses compute resources to back out trajectories for visualization and therefore slows compute time. It is not suggested that this parameter is set to true during a deployed use, but is a useful debug instrument while tuning the system, but use sparingly. Visualizating 2000 batches @ 56 points at 30 hz is alot.

The most common parameters you might want to start off changing are the velocity profiles (vx_max, vx_min, wz_max, and vy_max if holonomic) and the motion_model to correspond to your vehicle. Its wise to consider the prune_distance of the path plan in proportion to your maximum velocity and prediction horizon. The only deeper parameter that will likely need to be adjusted for your particular settings is the Obstacle critics' repulsion_weight since the tuning of this is proprtional to your inflation layer's radius. Higher radii should correspond to reduced repulsion_weight due to the penalty formation (e.g. inflation_radius - min_dist_to_obstacle). If this penalty is too high, the robot will slow significantly when entering cost-space from non-cost space or jitter in narrow corridors. It is noteworthy, but likely not necessary to be changed, that the Obstacle critic may use the full footprint information if consider_footprint = true, though comes at an increased compute cost.

Prediction Horizon, Costmap Sizing, and Offsets

As this is a predictive planner, there is some relationship between maximum speed, prediction times, and costmap size that users should keep in mind while tuning for their application. If a controller server costmap is set to 3.0m in size, that means that with the robot in the center, there is 1.5m of information on either side of the robot. When your prediction horizon (time_steps * model_dt) at maximum speed (vx_max) is larger than this, then your robot will be artifically limited in its maximum speeds and behavior by the costmap limitation. For example, if you predict forward 3 seconds (60 steps @ 0.05s per step) at 0.5m/s maximum speed, the minimum required costmap radius is 1.5m - or 3m total width.

The same applies to the Path Follow and Align offsets from furthest. In the same example if the furthest point we can consider is already at the edge of the costmap, then further offsets are thresholded because they're unusable. So its important while selecting these parameters to make sure that the theoretical offsets can exist on the costmap settings selected with the maximum prediction horizon and velocities desired.

The Path Follow critic cannot drive velocities greater than the projectable distance of that velocity on the available path on the rolling costmap. The Path Align critic offset_from_furthest represents the number of path points a trajectory passes through while tracking the path. If this is set either absurdly low (e.g. 5) it can trigger when a robot is simply trying to start path tracking causing some suboptimal behaviors and local minima while starting a task. If it is set absurdly high (e.g. 50) relative to the path resolution and costmap size, then the critic may never trigger or only do so when at full-speed. A balance here is wise. A selection of this value to be ~30% of the maximum velocity distance projected is good (e.g. if a planner produces points every 2.5cm, 60 can fit on the 1.5m local costmap radius. If the max speed is 0.5m/s with a 3s prediction time, then 20 points represents 33% of the maximum speed projected over the prediction horizon onto the path). When in doubt, prediction_horizon_s * max_speed / path_resolution / 3.0 is a good baseline.

Obstacle, Inflation Layer, and Path Following

There also exists a relationship between the costmap configurations and the Obstacle critic configurations. If the Obstacle critic is not well tuned with the costmap parameters (inflation radius, scale) it can cause the robot to wobble significantly as it attempts to take finitely lower-cost trajectories with a slightly lower cost in exchange for jerky motion. It may also perform awkward maneuvors when in free-space to try to maximize time in a small pocket of 0-cost over a more natural motion which involves moving into some low-costed region. Finally, it may generally refuse to go into costed space at all when starting in a free 0-cost space if the gain is set disproportionately higher than the Path Follow scoring to encourage the robot to move along the path. This is due to the critic cost of staying in free space becoming more attractive than entering even lightly costed space in exchange for progression along the task.

Thus, care should be taken to select weights of the obstacle critic in conjunction with the costmap inflation radius and scale so that a robot does not have such issues. How I (Steve, your friendly neighborhood navigator) tuned this was to first create the appropriate obstacle critic behavior desirable in conjunction with the inflation layer parameters. Its worth noting that the Obstacle critic converts the cost into a distance from obstacles, so the nature of the distribution of costs in the inflation isn't overly significant. However, the inflation radius and the scale will define the cost at the end of the distribution where free-space meets the lowest cost value within the radius. So testing for quality behavior when going over that threshold should be considered.

As you increase or decrease your weights on the Obstacle, you may notice the aforementioned behaviors (e.g. won't overcome free to non-free threshold). To overcome them, increase the FollowPath critic cost to increase the desire for the trajectory planner to continue moving towards the goal. Make sure to not overshoot this though, keep them balanced. A desirable outcome is smooth motion roughly in the center of spaces without significant close interactions with obstacles. It shouldn't be perfectly following a path yet nor should the output velocity be wobbling jaggedly.

Once you have your obstacle avoidance behavior tuned and matched with an appropriate path following penalty, tune the Path Align critic to align with the path. If you design exact-path-alignment behavior, its possible to skip the obstacle critic step as highly tuning the system to follow the path will give it less ability to deviate to avoid obstacles (though it'll slow and stop). Tuning the critic weight for the Obstacle critic high will do the job to avoid near-collisions but the repulsion weight is largely unnecessary to you. For others wanting more dynamic behavior, it can be beneficial to slowly lower the weight on the obstacle critic to give the path alignment critic some more room to work. If your path was generated with a cost-aware planner (like all provided by Nav2) and providing paths sufficiently far from obstacles for your satesfaction, the impact of a slightly reduced Obstacle critic with a Path Alignment critic will do you well. Not over-weighting the path align critic will allow the robot to deviate from the path to get around dynamic obstacles in the scene or other obstacles not previous considered duing path planning. It is subjective as to the best behavior for your application, but it has been shown that MPPI can be an exact path tracker and/or avoid dynamic obstacles very fluidly and everywhere in between. The defaults provided are in the generally right regime for a balanced initial trade-off.