Learning Task-Invariant Properties Via Dreamer: Enabling Efficient Policy Transfer for Quadruped Robots
Junyang Liang, Yuxuan Liu, Yanbin Chang, Junfan Lin, JunKai JI, Hui Li, Changxin Huang, Jianqiang Li
AI summary
Problem
Sim-to-real transfer for quadruped robots struggles with dynamics discrepancies that cause performance drops, while existing methods rely on costly real-world fine-tuning or manual feature design prone to catastrophic forgetting.
Approach
DreamTIP leverages an LLM to extract robust physical properties like contact stability and terrain clearance, integrating them as auxiliary targets in a Dreamer world model alongside a mixed replay buffer and regularization for rapid real-world adaptation.
Key results
- LLM-driven extraction of task-invariant properties reduces reliance on specific simulation dynamics
- Achieves 28.1% average performance improvement across eight simulated transfer tasks
- Reaches 100% success rate in real-world climbing tasks versus 10% for baselines
- Mitigates representation collapse and catastrophic forgetting during few-shot real-world adaptation
Why it matters
Provides a scalable, low-data sim-to-real transfer method for legged robots, advancing robust autonomous locomotion in complex real-world environments.
Abstract
Achieving quadruped robot locomotion across diverse and dynamic terrains presents significant challenges, primarily due to the discrepancies between simulation envi- ronments and real-world conditions. Traditional sim-to-real transfer methods often rely on manual feature design or costly real-world fine-tuning. To address these limitations, this paper proposes the DreamTIP framework, which incorporates Task- Invariant Properties learning within the Dreamer world model architecture to enhance sim-to-real transfer capabilities. Guided by large language models, DreamTIP identifies and leverages Task-Invariant Properties, such as contact stability and terrain clearance, which exhibit robustness to dynamic variations and strong transferability across tasks. These properties are integrated into the world model as auxiliary prediction targets, enabling the policy to learn representations that are insensitive to underlying dynamic changes. Furthermore, an efficient adap- tation strategy is designed, employing a mixed replay buffer and regularization constraints to rapidly calibrate to real-world dynamics while effectively mitigating representation collapse and catastrophic forgetting. Extensive experiments on complex terrains, including Stair, Climb, Tilt, and Crawl, demon- strate that DreamTIP significantly outperforms state-of-the- art baselines in both simulated and real-world environments. Our method achieves an average performance improvement of 28.1% across eight distinct simulated transfer tasks. In the real-world Climb task, the baseline method achieved only a 10% success rate, whereas our method attained a 100% success rate. These results indicate that incorporating Task-Invariant Properties into Dreamer learning offers a novel solution for achieving robust and transferable robot locomotion.