KD-Judge: A Knowledge-Driven Automated Judge Framework for Functional Fitness Movements on Edge Devices

A research team led by Shaibal Saha has introduced KD-Judge, a novel automated judging system designed specifically for functional fitness movements like those in CrossFit competitions. The framework addresses a critical gap in sports technology: while AI scoring systems exist, they typically lack the transparency and rule-grounded precision needed for official judging. KD-Judge's breakthrough comes from its two-stage pipeline that first converts unstructured, natural language rulebooks into machine-readable, executable representations using an LLM-based retrieval-augmented generation (RAG) and chain-of-thought reasoning approach.

Once rules are structured, the system employs deterministic rule-based judging combined with pose-guided kinematic reasoning to assess repetition validity and temporal boundaries at the individual rep level. This creates an auditable, transparent decision trail. For practical deployment, the team engineered a dual-strategy caching mechanism that dramatically reduces computational overhead on edge devices. Testing on the CFRep dataset showed the system operates faster than real-time (RTF < 1), with caching delivering up to 3.36x speedup for pre-recorded analysis and a remarkable 15.91x acceleration for live-streaming scenarios on the resource-constrained NVIDIA Jetson AGX Xavier platform.

The research, accepted at IEEE/ACM CHASE 2026, demonstrates that KD-Judge can reliably structure complex fitness rules and execute accurate, frame-by-frame movement analysis. This enables scalable, consistent judging that complements human officials, potentially reducing subjectivity and time pressures in competitive environments. The edge-optimized architecture means the system could be deployed in gyms, at competition venues, or even for personal training feedback without relying on cloud connectivity.