Rover PathPlanner 2025

This repository contains two implementations of the same autonomous terrain planner:

• a reference Python prototype that uses NumPy and Pillow for image driven A* path planning
• a Rust port that mirrors the Python package layout but adds profiling, tracing and better performance

Both versions share identical configuration defaults so you can compare behaviour, speed and visual output side by side.

Sample outputs

Python run (pathplanner/output/):

Rust run (rustfinder/output/):

Why this project

The original software department challenge (described in TASKS.md) asked us to guide a rover across a 100x100 height map, obeying a 30deg slope limit. The work evolved in three phases:

1. Python prototype (documented in IDEAS.md) that builds the image pipeline, cost-map and baseline A*.
2. Visual tooling: red-channel filtering, cost heat map, blocked cell overlays and animated GIF playback of the search.
3. Rust rewrite that keeps the same architecture, then layers in bidirectional A* with slope-aware cost scaling, sampling controls and end-point tolerance. Benchmarks in TIME.md show the Rust binary running roughly twice as fast as the Python script.

The repository keeps both code paths so the evolution is easy to audit and compare.

Project structure

pathplanner/         # Python package
  config.py          # shared configuration defaults
  finder/            # bidirectional A*, conversion helpers, tracing hooks
  image/             # synthetic map generation, filtering, animation
  main.py            # end-to-end pipeline entry point

rustfinder/          # Cargo project with mirrored modules
  src/
    config.rs        # same defaults as config.py
    finder/          # bidirectional A*, tracing structs
    image/           # image IO, GIF renderer
    main.rs          # Rust entry point with CLI flags

IDEAS.md             # design notes for each milestone
TASKS.md             # original challenge brief
TIME.md              # quick benchmark snapshot
README.md            # you are here

Features

• Synthetic terrain generation (Gaussian blurred RGB noise with red-channel extraction)
• Heightmap conversion from pixel intensity to metres, followed by gradient-aware cost weighting
• Bidirectional A* search with:
  • slope filtering on each edge
  • movement modes (4 or 8 directions)
  • configurable trace sampling to control animation size
  • goal tolerance radius (default 5px) so runs stop once any cell close to the target is reached
• Step-by-step tracing API that powers both console logging and GIF playback
• Identical configuration surface across Python and Rust implementations

Configuration

Key fields from config.py / src/config.rs:

Setting	Default	Description
IMAGE_SIZE	100	Generated map width/height in pixels
BLUR_SIZE	3.0	Gaussian blur radius for terrain smoothing
HEIGHT_MIN_M / HEIGHT_MAX_M	0 / 3	Height range used when converting the red channel
PIXEL_SIZE_M	0.1	Physical scale per pixel
MAX_SLOPE_DEG	30	Maximum climb angle
SLOPE_COST_WEIGHT	100	Additional traversal cost applied for steep slopes
REMAINING_DISTANCE_WEIGHT	5	Heuristic multiplier for A*
TRACE_SAMPLE_STRIDE	50	How many A* steps to skip between stored snapshots
END_TOLERANCE_PX	5	Radius around the goal that is considered “close enough”

Edit config.py to tweak the Python run. The Rust binary reads the same values from src/config.rs.

Running the Python pipeline

cd pathplanner
uv run ./main.py

Outputs:

• output/noisy.png – generated terrain
• output/red.png – red-channel heightmap
• output/path.png – final route overlay with blocked cells
• output/search.gif – search animation (only when trace samples are captured)

Console output lists timing (Search time, Total pipeline time) for quick iteration.

Running the Rust pipeline

cd rustfinder
cargo run . --trace --release

The Rust binary writes the same image and GIF artefacts into rustfinder/output/.

Comparing Python vs Rust

TIME.md holds a quick timing snapshot (Rust ~2.45s vs Python ~3.81s on the author’s machine). To repeat:

1. Run uv run ./main.py and note the printed Search time.
2. Run cargo run . --trace --release and note the reported timings.

Both runs use the same config defaults so the path, generated artefacts and GIF should match closely.