ARM: Advantage Reward Modeling for Long-Horizon Manipulation

2026-04-03

AI Summary

This work addresses the challenge of credit assignment in long-horizon robotic manipulation, where sparse rewards hinder learning and dense rewards are costly to annotate and ill-suited for non-monotonic behaviors. The authors propose Advantage-based Reward Modeling (ARM), a framework that replaces absolute progress evaluation with relative advantage comparisons between actions. ARM leverages low-cognitive-burden ternary human labels—indicating progress, regression, or stagnation—to enable consistent yet inexpensive supervision. The approach automatically generates rewards from both complete and fragmented demonstrations and integrates offline reinforcement learning with DAgger-style data utilization. Evaluated on a complex towel-folding task, ARM achieves a 99.4% success rate, substantially outperforming existing Vision-Language-Action (VLA) baselines while demonstrating exceptional data efficiency and training stability.

Abstract

Long-horizon robotic manipulation remains challenging for reinforcement learning (RL) because sparse rewards provide limited guidance for credit assignment. Practical policy improvement thus relies on richer intermediate supervision, such as dense progress rewards, which are costly to obtain and ill-suited to non-monotonic behaviors such as backtracking and recovery. To address this, we propose Advantage Reward Modeling (ARM), a framework that shifts from hard-to-quantify absolute progress to estimating relative advantage. We introduce a cost-effective tri-state labeling strategy -- Progressive, Regressive, and Stagnant -- that reduces human cognitive overhead while ensuring high cross-annotator consistency. By training on these intuitive signals, ARM enables automated progress annotation for both complete demonstrations and fragmented DAgger-style data. Integrating ARM into an offline RL pipeline allows for adaptive action-reward reweighting, effectively filtering suboptimal samples. Our approach achieves a 99.4% success rate on a challenging long-horizon towel-folding task, demonstrating improved stability and data efficiency over current VLA baselines with near-zero human intervention during policy training.

Problem

• long-horizon manipulation
• sparse rewards
• credit assignment
• dense progress rewards
• non-monotonic behaviors

Innovation

• Advantage Reward Modeling
• long-horizon manipulation
• tri-state labeling
• offline reinforcement learning
• progress estimation