Beyond Pixels: Learning Invariant Rewards for Real-World Robotics From a Few Demonstrations

TL;DR

This paper introduces a method for learning invariant, symbolic reward functions from minimal demonstrations, enabling better generalization and zero-shot transfer in real-world robotic manipulation tasks.

Contribution

It proposes a novel framework that shifts from visual feature fitting to discovering behavioral invariants, improving reward generalization across diverse task variants.

Findings

Achieves stronger process alignment and policy ranking on Meta-World and Franka tasks.

Demonstrates zero-shot generalization to position, viewpoint, and object variations.

Accelerates downstream policy learning with learned invariant rewards.

Abstract

Designing reward functions that generalize beyond controlled laboratory settings remains a fundamental challenge in reinforcement learning for robotics. In open-world manipulation problems, a single task can appear in numerous variants through different object instances, positions, and camera viewpoints. Recent vision-based reward models tend to memorize specific pixel distributions and fail to generalize beyond their training conditions. To address this, we propose a framework that learns invariant symbolic reward functions from as few as five demonstrations. The insight is to shift from visual feature-fitting to the discovery of behavioral invariants: task-level properties that remain constant across diverse visual instantiations. The framework has two coupled components: a structural reward formulation that encodes task-level strategies and physical constraints while preserving optimal policy invariance, and a hybrid symbolic-numerical procedure that distills these invariants from demonstrations without online interaction. Experiments on eight Meta-World tasks and three Franka manipulation tasks demonstrate that our method achieves stronger process alignment and policy rollout ranking abilities compared to baselines, accelerating downstream policy learning. Three real-world out-of-distribution experiments further show that the same learned reward generalizes zero-shot to position, viewpoint, and object variations, enabling a single reward representation to be reused across diverse task variants in practice.