beam-tracking-ml

Design and refactor beam tracking ML/RL pipelines (CSI teacher vs RSRP student), enforce shape contracts, and produce inference-safe models.

⭐0stars🔀0forks📁View on GitHub🕐Updated Jan 5, 2026

When & Why to Use This Skill

This Claude skill specializes in designing and refactoring machine learning (ML) and reinforcement learning (RL) pipelines specifically for beam tracking in wireless communications. It facilitates the transition from RL architectural diagrams to modular code, ensures rigorous tensor shape validation for inference safety, and implements advanced teacher-student distillation workflows to optimize models for real-time deployment.

Use Cases

  • Architectural Translation: Converting complex RL framework diagrams into modular, maintainable Python components for wireless signal processing.
  • Code Refactoring: Streamlining monolithic scripts (like Sionna-based tracking) into organized directory structures for models, training scripts, and schema interfaces.
  • Knowledge Distillation: Implementing teacher-student workflows to transfer knowledge from CSI-heavy offline models to lightweight, RSRP-based online policies for edge deployment.
  • Shape Contract Enforcement: Defining and testing strict tensor dimensions (e.g., Batch, N_BEAMS) across the pipeline to ensure robust, error-free model inference.
  • Schema Design: Creating standardized observation and action schemas to ensure compatibility between ML models and wireless xApp runtime environments.
namebeam-tracking-ml
descriptionDesign and refactor beam tracking ML/RL pipelines (CSI teacher vs RSRP student), enforce shape contracts, and produce inference-safe models.
allowed-toolsRead, Grep, Glob, Bash

Beam Tracking ML Skill

Use this Skill when:

  • translating the RL架構 diagram into code
  • refactoring sionna_beam_tracking_v2.py ideas into modular components
  • designing observation/action schemas

Guardrails

  • Always define and test shapes (B,N_BEAMS) etc.
  • Keep student (online) policy lightweight and deterministic.
  • Treat CSI-heavy path as offline only unless we explicitly design compression.

Where to put code

  • Models: beam_tracking/model/
  • Training scripts: scripts/ (do not bloat runtime xApp)
  • Interfaces: beam_tracking/schemas.py

Suggested distillation workflow

  1. Train teacher on CSI dataset (offline).
  2. Run teacher over same trajectories, log action distributions.
  3. Train student to match teacher (KL divergence).
  4. Optionally fine-tune student with small online data.