Plan: I-MPPI (Informative MPPI) Development

This plan outlines the steps for developing the i_mppi controller variant in the cuda-mppi library, based on the “Hierarchical Informative MPPI” framework.

Goal

Implement IMPPIController, a controller that integrates:

1. Biased Sampling: A mixture distribution of samples centered around the previous solution and a global reference trajectory.
2. Information Cost: A cost term based on Fast Shannon Mutual Information (FSMI) to reward exploration.

Steps

1. Research and Definition

• Analyze /home/riccardo/phd/notes/delft/i_mppi_note.qmd.
• Define strategy:
  • Biased Sampling: Implemented by shifting the mean of the noise buffer for a subset of samples. \(u = u_{nom} + \epsilon\), where \(\epsilon\) is shifted such that effectively \(u \sim \mathcal{N}(u_{ref}, \Sigma)\).
  • FSMI Cost: A custom Cost functor that performs raycasting on a GPU-resident map (Occupancy Grid for MVP).

2. Infrastructure Setup

• Create development branch feat/i-mppi.
• Create development folder i_mppi/.
• Initialize include/mppi/controllers/i_mppi.cuh.

3. Implementation

• Config: Add alpha (bias ratio), lambda_info (info gain weight) to MPPIConfig (or a subclass).
• Map Representation: Create a simple OccupancyGrid CUDA struct in include/mppi/core/map.cuh (new file) to support raycasting.
• FSMI Cost: Implement FSMICost in include/mppi/instantiations/fsmi_cost.cuh.
  • Implement raycast and compute_info_gain device functions.
• Controller: Implement IMPPIController in include/mppi/controllers/i_mppi.cuh.
  • Manage u_ref (reference trajectory).
  • Implement shift_noise_means kernel/function to apply bias.

4. Testing & Verification

• Create a standalone test in src/i_mppi_test.cu.
  • Setup a mock map (e.g., a wall with a hole or a simple room).
  • Verify that the controller is attracted to unknown areas.
• Benchmark performance of the raycasting kernel. (Implicitly verified functionality via test run)
• Conduct extensive testing campaign and generate documentation images (src/i_mppi_sim.cu).

5. Integration

• Expose new classes to Python bindings if needed (low priority for CLI task, focus on C++ core).
• Cleanup and Documentation update.

Progress Tracking

• 2026-02-05: Updated plan with specific I-MPPI details.
• 2026-02-05: Implemented IMPPIController, FSMICost, OccupancyGrid and verified with i_mppi_test.
• 2026-02-05: Conducted extensive testing campaign, generated comparison plots and updated documentation.