Learning to Coach (L2C): an RL pipeline combining adaptive blending, a skill-change PFA, and a surrogate reward

Design · All designs
Evidence · 3 claims (3 for) · 1 study (1 causal), q3 · 1 of 1 report an effect size · 3 claims rest on one study

Description

L2C is the article's trainable coaching framework built on the Coaching Game formalism. Its contributions are "threefold: (i) a closed-loop blending rule that adapts the assistance levelλto the coach's observations of both the physical environment and the learner behavior; (ii) a reformu-lation of theCoaching Gameto a single-agent POMDP, solvable via RL, through a probabilistic model of the coach's causal influence on learner skill; and (iii) a tractable surrogate reward that is provably consistent with the counterfactual V oI objective." A hybrid probabilistic finite-state automaton models skill change on success and failure events with sigmoid rates over skill; the coach policy is trained with PPO against skill improvement and deployed with Bayesian runtime skill inference and multimodal feedback.

Design Implications

Context

Requirements

• A pretrained expert policy to anchor the blending rule
• A learner model expressible as a skill-parameterized policy class
• Privileged access to learner skill during simulation training

Constraints

• The Boltzmann learner model is one tractable choice; the article states accurately modeling skill-conditioned learner behaviors remains an open problem
• Verbal and visual cues are predefined and triggered by fixed rules; only assistance modulation of physical control is learned

Target Learners

• human learners practicing motor skills such as FPV drone racing, mostly novices

Learning Goals

• accelerated development of independent motor-skill competence, measured by unassisted lap time and failure count

Claims

• L2C Within Subject Lap Time Failure Gains [+M]
• L2C Outperforms Rbf Mia Baselines [+M]
• L2C Adaptive Assistance By Skill And Context [+M]

Related Designs

Examples

-

Key Sources

• Wei Wang, Enlin Gu, Antonio Loquercio, Haimin Hu, Rahul Mangharam. (2026). AI Coaching for Accelerating Human Skill Development with Reinforcement Learning. arXiv preprint. https://arxiv.org/abs/2606.25337

Page metadata

Field	Value
type	design
id	l2c-rl-coaching-framework
title	Learning to Coach (L2C): an RL pipeline combining adaptive blending, a skill-change PFA, and a surrogate reward
description	L2C is the article's trainable coaching framework built on the Coaching Game formalism.
status	review
generated	by process:settle-candidates on 2026-10-10
trust	unverified — no one has confirmed this page's content
sources	1 — wei-wang-2026

Frontmatter is the page's machine-readable record — see the Schema & Guide.