![[2502.00835] CAIMAN: Causal Action Influence Detection for Sample-efficient Loco-manipulation](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2502.00835] CAIMAN: Causal Action Influence Detection for Sample-efficient Loco-manipulation
Summary
The paper introduces CAIMAN, a reinforcement learning framework designed to enhance legged robots' capabilities in non-prehensile loco-manipulation by utilizing causal action influence for efficient skill acquisition.
Why It Matters
As robotics continues to evolve, developing efficient methods for robots to interact with their environments is crucial. CAIMAN addresses the challenge of enabling legged robots to perform complex tasks with minimal guidance, which could significantly improve their practical applications in real-world scenarios.
Key Takeaways
- CAIMAN promotes sample-efficient learning for legged robots.
- The framework uses causal action influence as intrinsic motivation.
- It combines hierarchical control strategies for improved adaptability.
- Empirical results show superior performance in simulations and real-world applications.
- The approach minimizes the need for extensive task-specific reward shaping.
Computer Science > Robotics arXiv:2502.00835 (cs) [Submitted on 2 Feb 2025 (v1), last revised 20 Feb 2026 (this version, v3)] Title:CAIMAN: Causal Action Influence Detection for Sample-efficient Loco-manipulation Authors:Yuanchen Yuan, Jin Cheng, Núria Armengol Urpí, Stelian Coros View a PDF of the paper titled CAIMAN: Causal Action Influence Detection for Sample-efficient Loco-manipulation, by Yuanchen Yuan and 3 other authors View PDF HTML (experimental) Abstract:Enabling legged robots to perform non-prehensile loco-manipulation is crucial for enhancing their versatility. Learning behaviors such as whole-body object pushing often requires sophisticated planning strategies or extensive task-specific reward shaping, especially in unstructured environments. In this work, we present CAIMAN, a practical reinforcement learning framework that encourages the agent to gain control over other entities in the environment. CAIMAN leverages causal action influence as an intrinsic motivation objective, allowing legged robots to efficiently acquire object pushing skills even under sparse task rewards. We employ a hierarchical control strategy, combining a low-level locomotion module with a high-level policy that generates task-relevant velocity commands and is trained to maximize the intrinsic reward. To estimate causal action influence, we learn the dynamics of the environment by integrating a kinematic prior with data collected during training. We empirically demonstrate CAIMAN's sup...