![[2602.16863] SimToolReal: An Object-Centric Policy for Zero-Shot Dexterous Tool Manipulation](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.16863] SimToolReal: An Object-Centric Policy for Zero-Shot Dexterous Tool Manipulation
Summary
SimToolReal presents a novel approach to zero-shot dexterous tool manipulation using an object-centric policy, enhancing robotic capabilities without task-specific training.
Why It Matters
This research addresses the challenges of tool manipulation in robotics, a key area for advancing automation. By enabling robots to adapt to various tools without extensive retraining, it opens pathways for more versatile applications in real-world scenarios, potentially transforming industries reliant on robotic assistance.
Key Takeaways
- SimToolReal uses procedural generation for diverse tool-like objects, enhancing generalization in robotic manipulation.
- The approach outperforms traditional methods by 37%, demonstrating significant improvements in dexterous manipulation.
- Zero-shot performance is achieved across 120 real-world rollouts, showcasing the model's adaptability to various tasks and tools.
- The research reduces the need for extensive engineering efforts in sim-to-real reinforcement learning.
- This advancement could lead to more efficient robotic systems in everyday applications.
Computer Science > Robotics arXiv:2602.16863 (cs) [Submitted on 18 Feb 2026] Title:SimToolReal: An Object-Centric Policy for Zero-Shot Dexterous Tool Manipulation Authors:Kushal Kedia, Tyler Ga Wei Lum, Jeannette Bohg, C. Karen Liu View a PDF of the paper titled SimToolReal: An Object-Centric Policy for Zero-Shot Dexterous Tool Manipulation, by Kushal Kedia and 3 other authors View PDF HTML (experimental) Abstract:The ability to manipulate tools significantly expands the set of tasks a robot can perform. Yet, tool manipulation represents a challenging class of dexterity, requiring grasping thin objects, in-hand object rotations, and forceful interactions. Since collecting teleoperation data for these behaviors is challenging, sim-to-real reinforcement learning (RL) is a promising alternative. However, prior approaches typically require substantial engineering effort to model objects and tune reward functions for each task. In this work, we propose SimToolReal, taking a step towards generalizing sim-to-real RL policies for tool manipulation. Instead of focusing on a single object and task, we procedurally generate a large variety of tool-like object primitives in simulation and train a single RL policy with the universal goal of manipulating each object to random goal poses. This approach enables SimToolReal to perform general dexterous tool manipulation at test-time without any object or task-specific training. We demonstrate that SimToolReal outperforms prior retargeting ...