A neural network for climbing route generation

Hribar, Samo (avtor), Kristan, Matej (mentor), Oblak, Polona (komentor)

Grading and setting of climbing routes is a challenging task, even for experienced climbers: climbing walls vary in angles, each has a variable number of holds, each of them having unique shapes and spatial arrangement, with climbers routinely disagreeing on difficulty ratings by a full grade. While several recent works have proposed automatic route scoring and generation, they are tailored to a specific board. Moreover, there is a lack of evaluation of the competing methods on an extensive common benchmark, which hinders progress. To address these shortcomings, we present BERNARD, a system of two transformer-based neural networks that is capable of scoring and generating routes across eight different board layouts from three commercial board families. We propose the Boulder Evaluation and Rating Network (BERN), which predicts the difficulty, its uncertainty and route quality. Additionally, we propose the Neural Attention-based Route Diffuser (NARD), which formulates generation as flow matching over hold activations, conditioned on route difficulty, quality rating, angle and, as the first generator, the hand-match prohibition. As a secondary contribution, we propose BERNOT, the currently largest curated benchmark with over 477,000 train, 68,000 validation and 136,000 test routes, spanning eight boards with a novel route scoring and route setting evaluation protocol. Extensive evaluation on BERNOT shows that BERN outperforms other state-of-the-art scorers in MAE by 20.1% on average and on all eight boards, and NARD outperforms the current best state-of-the-art generator by 4.3% in MAE and 2.4 points in accuracy ±1 on average.

• Jezik: Angleški jezik
• Ključne besede: Neural networks, Climbing, Generating boulders, Automatic boulder grading
• Vrsta gradiva: Magistrsko delo/naloga
• Organizacija: FRI - Fakulteta za računalništvo in informatiko
• Leto izida: 2026
• PID: 20.500.12556/RUL-190741
• Datum objave v RUL: 02.11.2026

Sekundarni jezik

• Jezik: Slovenski jezik
• Naslov: Nevronska mreža za generiranje plezalnih smeri

Izvleček

Ocenjevanje in postavljanje balvanov je zahtevna naloga celo za izkušene plezalce: plezalne stene imajo različne naklone in različno število, obliko ter razporeditev oprimkov, plezalci pa pogosto različno ocenijo težavnost smeri. Kljub obstoječim metodam, ki omogočajo samodejno ocenjevanje in generiranje balvanov, so le-te prilagojene posamezni plezalni steni. Poleg tega ne obstaja enotna urejena zbirka balvanov, ki bi omogočala primerjavo različnih metod. V našem delu predstavimo BERNARD, sistem dveh transformerskih nevronskih mrež, ki ocenjuje in generira balvane na osmih različnih stenah treh proizvajalcev. Predlagamo mrežo BERN (Boulder Evaluation and Rating Network), ki napove težavnost, negotovost težavnosti in kakovost balvana. Predlagamo tudi mrežo NARD (Neural Attention-based Route Diffuser), ki generiranje smeri rešuje s pristopom izravnalnega toka (angl., flow matching) nad aktivacijami oprimkov, pogojenega z naklonom stene, želeno težavnostjo in kakovostjo balvana, hkrati pa je tudi prvi generator, ki kot pogoj podpira tudi prepoved preprijemanja oprimkov. Kot sekundarni prispevek predstavimo BERNOT, trenutno največjo prečiščeno zbirko z več kot 477.000 učnimi, 68.000 validacijskimi in 136.000 testnimi balvani na osmih stenah, z novim protokolom za vrednotenje ocenjevanja in postavljanja balvanov. Obsežno vrednotenje na zbirki BERNOT pokaže, da BERN prekaša najboljše obstoječe metode za ocenjevanje, v povprečju za 20,1% po meri MAE na vseh osmih stenah, NARD pa prekaša najboljši obstoječi generator v povprečju za 4,3% MAE in za 2,4 enote točnosti v intervalu ±1.

• Ključne besede: Nevronske mreže, Plezanje, Generiranje balvanov, Samodejno ocenjevanje balvanov