Skip to content

Mode Physical (P) — Judging

2025.1 rubric summary

See Competition → Rules & Scoring for the current 2025.1 weights and the optional Edge‑AI Bonus. This page retains 2024 reference categories for context.

Algorithm-agnostic

Credit is based on outcomes and evidence; architecture choice (mechanical/control/ML/LLM) is not scored directly, provided constraints and safety are met.

Innovation Uplift

Judges may award top-of-band within categories for documented novel contributions (mechanical design, sensing, safety engineering, UI/HF overlays, calibration tooling, data quality pipelines, model compression/efficiency) that improve outcomes, safety, or clarity. This does not exceed category maxima.

Small multimodal LLMs encouraged

  • Expectation: focus your effort on generating/fine-tuning data and optimizing compact multimodal models.
  • Inference‑only, offline, JSON outputs (e.g., Explain‑This‑Event or Rig‑State Textualizer).
  • Small‑model, edge‑deployable focus: teams may prototype with larger cloud‑hosted models during development, but final deliverables and performance are evaluated using locally run or on‑rig small models under limited inference hardware.
  • Suggested caps: ≤2 vCPU or ≤20 W GPU, ≤4–8 GB RAM (non‑LLM) and ≤8 vCPU or ≤1,150 W GPU, ≤16–64 GB RAM (LLM); ≤15 min total inference runtime. Rig compute otherwise open. All AI runs must be offline during judging; no PLC/actuator writes.
  • Custom task option: teams may propose a custom offline JSON task with objective ground truth and the same interface for pre‑approval (recommended deadline: 2026‑02‑01). Approved tasks use the same AI sub‑rubric and scale to +10.

Judging spans both the design phase and the final test. The 2024 guidelines define the categories below; detailed weightings are in the source document.

Phase I (design) categories:

  • Safety
  • Mobility of rig (rig up/move/rig down)
  • Design considerations and lessons learned (analysis, concepts, development)
  • Mechanical design and functionality/versatility (problem/goal, criteria, feasibility)
  • Simulation/model/algorithm
  • Control scheme

Phase II (build/test) categories:

  • Creative ability; engineering skills; construction quality; cost control
  • Performance (e.g., ROP, MSE, landing bit, inclination; proximity to target X/Y and depth; Go/No-Go flexible “casing”; verticality, tortuosity, caliper)
  • Data (handling, visualization, comparison with judges’ logs)
  • Downhole sensor data used in the control algorithm
  • Intangibles and pass/fail checks

Note: Bonus points may be added or subtracted at judges’ discretion. See 2024/group-b.md for specific weights and examples extracted from the PDF.