AIAny
Icon for item

Progress Reward Modeling for Robotic Learning: A Comprehensive Survey

Provides a unified survey of progress-reward modeling for robotic learning, detailing interfaces, modeling techniques, and evaluation practices. Organizes the literature into three perspectives—interface, model internals, and data/benchmarks—and highlights limitations and open problems. Useful for researchers designing rewards for long-horizon or sparse-reward robotic tasks.

Introduction

Why this matters

Robotic learning commonly suffers from sparse terminal rewards: a success/failure signal does not reveal whether current behavior is improving or regressing. This survey argues that crafting and evaluating progress rewards—signals that indicate incremental task progress—is essential to make learning sample-efficient and reliable in long-horizon, compositional robotic tasks.

Key Findings
  • A unifying interface perspective: progress models differ in what they observe (eg. state, demonstration, goal specification) and what they output (scalar progress, stepwise rewards, or ranking). So what: reconciling interfaces clarifies when a method transfers across task formulations and what additional supervision is required.

  • Diverse modeling families: methods include handcrafted heuristics, supervised regressors trained on demonstrations, contrastive/ranking objectives, and learned progress estimators that leverage representation learning. So what: each family trades off data needs, generalization, and robustness; newer learned estimators scale but demand careful supervision or proxies.

  • Evaluation & data mismatch: benchmarks and supervision sources vary widely (synthetic environments, scripted demonstrations, human labels), which complicates comparisons. So what: the survey emphasizes standardized protocols and metrics to judge whether a progress signal actually speeds up learning versus just correlating with success.

Who It's For and Trade-offs

Great fit if you are a robotics or RL researcher designing reward functions or progress estimators for long-horizon, sparse-reward tasks, or if you need a taxonomy to select supervision sources and evaluation protocols. Look elsewhere if you only need low-level control primitives or purely model-free baselines where terminal rewards suffice.

Trade-offs the survey highlights: learned progress signals can improve exploration but require representative data and careful validation; simple heuristics are cheap but brittle across task variants; evaluation must measure downstream learning improvement, not just prediction accuracy.

Methods and organization

The paper structures the field into three connected steps: (1) interface—what inputs and outputs a progress model exposes, (2) interior—how models are constructed and trained (objectives, architectures, supervision types), and (3) data & benchmarks—how progress supervision is collected and how methods are evaluated. It closes with limitations (data scarcity, transferability, robust evaluation) and concrete research directions for reproducible comparisons and deployment in real robots.

Information

  • Websitearxiv.org
  • AuthorsJianshu Zhang, Keliang Wu, Haoran Lu, Anbang Liu, Ce Zhang, Weijie Yin, Chengxuan Qian, Xiyuan Yang, Zhenyu Pan, Guo Ye
  • Published date2026/07/22

More Items

Bridges the proprietary-to-open-source gap in agentic search by converting multi-step retrieval and reasoning traces into a structured, style-normalized JSON protocol and using it for joint distillation + RL. Produces denser supervision that improves student success rates while reducing style drift.

A PyTorch-native training framework for agentic reinforcement learning research that keeps researcher-facing code compact and editable. Uses an asynchronous loop to train multimodal and mixture-of-experts policies while never training on tokens the agent didn't generate; matches Megatron-style stacks under a comparable protocol and ships recipes and containers on GitHub.

Presents Skill Self-Play (Skill-SP), a co-evolutionary training loop where a proposer, solver, and dynamic skill controller generate, solve, and verify tasks conditioned on reusable skills — balancing verifiable execution with open-ended task diversity to boost LLM tool-use and reasoning.