Bratmobile Documentation

The purpose of this library is to showcase a framework for multi-step ahead planning using pure input control (see Braitenberg vehicles). The navigation problem is broken down into several unique closed-loop input controllers, called Tasks. Each Task produces a unique control behaviour (go straight, turn left/right 90 degrees) in response to a Disturbance (an obstacle or target), which determines a Task's duration. A supervising module, called the Configurator, can simulate sequences of Tasks in game engine Box2D, retain their outcomes in a cognitive map, which can be searched to extract plans. The physics simulation represents the robot's Core Knowledge (Spelke, 2007).

Features:

• Completely on-the-fly, instantaneous construction of an abstract cognitive map of the environment, represented as a tree of tasks
• Hybrid state-space representation with flexible discretization
• Use of an attention window (simulated distal sensor) to define the scope of Tasks
• Mapless navigation: robot is ready to go as is, no need for global sensors
• Causal reasoning on the naive robot
• Plans represents set of instructions, not trajectories: no solvers required, completely closed-loop and object oriented

Work in progress

• Closed-loop Task execution with goal-directed behaviour (driving towards target)
• Debugging plan checking and recycling
• Adaptive thresholding of acceptable noise in state observations

Documentation

Available at https://glafratta.github.io/bratmobile/

Publications

Giulia Lafratta, Bernd Porr, Christopher Chandler, Alice Miller; Closed-Loop Multistep Planning. Neural Computation 2025; 37 (7): 1288–1319. doi: https://doi.org/10.1162/neco_a_01761

Hardware

The indoor robot is equipped with

• 360 Parallax Continuous Rotation Servo motors (see here for wiring)
• A1 SLAMTEC LIDAR (see here for wiring)
• Raspberry Pi model 3b+

Prerequisites

Development packages

• G++ compiler
• CMake
• PiGPIO library
• OpenCV
• Boost
• XOrg
• LibGLU1
• Gtest

sudo apt install g++ cmake libpigpio-dev libopencv-dev libboost-all-dev xorg-dev libglu1-mesa-dev libgtest-dev

Compile from source

• LIDAR API
• Motors API
• Cpp Timer
• Box2D v2.4.1 ** if not installed automatically, go to box2d/build and run sudo make install

Build

cd CloCK
sh build.sh

Run

Navigation demo (Raspberry Pi)

• sudo ./targetless : this program demonstrates planning over a 1m distance horizon for a control goal that is not a target location but rather an objective to drive straight for the longest time with the least amount of disturbances
• sudo ./target: this program (under construction) demonstrates target seeking behaviour, where the target is imaginary and located at x=1.0m, y=0m. Run with options 0 [custom-stepDistance]: for turning debug options off. In debug mode, LIDAR coordinates, Box2D objects and robot trajectories are dumped into the /tmp folder. The stepDistance is the maximum distance covered by a single task, 1.0m by default.

Only planning and execution (tracking using dead reckoning) demonstrated

WARNING! Due to the point cloud clustering algorithm used (Partition), an obstacle such as a cul-de-sac will be detected as a solid box, so take into account that representation of concave objects may be inaccurate.

Simulation (x86 architecture)

• test/simulations/target_sim [folder_with_scans/] [bool: is_real_time] : runs a simulation of a planning and plan recycling scenario (not fully debugged). Args:
  • folder_with_scans : a folder containing 2D LIDAR scans
  • is_real_time : a boolean flag used to determine whether to use a timer to ensure that LIDAR callbacks are called every 200ms and motor callbacks every 100 (default=1)

Unit tests (x86 architecture)

run make test