QDTraj: Exploration of Diverse Trajectory Primitives for Articulated Objects Robotic Manipulation

Technical Summary: QDTraj: Exploration of Diverse Trajectory Primitives for Articulated Objects Robotic Manipulation

1. Problem Statement

Robotic manipulation in unstructured, domestic environments is significantly challenged by the presence of articulated objects (e.g., ovens, drawers, faucets). Current methods for acquiring expert data for these tasks face three primary hurdles:

• Data Acquisition Difficulty: Real-world teleoperation is expensive and slow, while simulated data often suffers from the "sim-to-real" gap.
• Lack of Diversity: Most existing methods focus on finding a single optimal trajectory. However, in real-world scenarios, a robot needs a repertoire of diverse solutions to adapt to unexpected obstacles, changing constraints, or unreachable grasp points.
• Task Complexity: Manipulating articulated objects involves complex, contact-rich interactions that are difficult to model with simple, hard-coded motion primitives.

2. Methodology

The authors propose QDTraj, a "plug-and-play" framework that uses Quality-Diversity (QD) algorithms to automatically generate a diverse set of high-performing low-level trajectory primitives in simulation.

A. Core Algorithm: MAP-Elites
Instead of searching for one optimal solution, QDTraj uses the MAP-Elites algorithm to populate a multi-dimensional archive.

• Genotype: Each individual is defined by a 6-DoF end-effector starting frame (3D position and unit quaternion orientation).
• Behavioral Descriptor: The archive is organized into a 3D grid based on the Cartesian starting position of the grasp.
• Selection & Mutation: The algorithm uses a selection-mutation loop where successful individuals are cloned and mutated (adding bounded noise to position and orientation) to explore the solution space.

B. Adaptive Interaction Policy
To bridge the gap between grasping and manipulation, the authors introduce an Adaptive Interaction Strategy:

1. Grasp: The robot reaches the starting pose and performs a force-closure grasp.
2. Compliant Control: The robot switches to a compliant mode (low joint gains) to absorb modeling errors, while the specific object joint of interest is actuated with high gains to move from an initial to a target state.
3. Evaluation (Fitness): The fitness score is determined by how much of the required joint displacement the robot successfully achieves before losing contact.

C. Key Contributions

• QD-based Trajectory Generation: The first application of Quality-Diversity algorithms specifically for generating trajectory primitives for articulated objects.
• Adaptive Interaction Action Space: A novel interaction strategy that moves away from hard-coded Cartesian directions (like "push" or "pull") toward a more flexible, physics-grounded compliant control approach.
• Plug-and-Play Framework: A method that requires only the object's URDF and the task definition to generate a massive library of trajectories.

4. Results and Evaluation

The method was validated through extensive simulation (using the Genesis engine) and real-world deployment on a Franka Emika FR3 robot.

• Superior Diversity: QDTraj generated at least 5 times more diverse trajectories for hinge and slider tasks compared to baseline methods (Where2Act and VAT-Mart).
• Quantitative Performance: In comparative studies, the combination of the "MAP-Elite-Explore" selection strategy and the "Adaptive-AS" interaction policy significantly outperformed all other configurations. For hinge joints, it increased the number of successful trajectories from 27 (baseline) to 2,944.
• Generalization: The method was tested on 30 different articulations from the PartNet-Mobility dataset, yielding an average of 704 different trajectories per task, demonstrating strong zero-shot generalization to new object geometries.
• Sim-to-Real Success: The authors demonstrated that trajectories generated in simulation—including non-intuitive ones (e.g., grasping a door above the handle because the handle was unreachable)—successfully transferred to the real robot.

5. Significance

QDTraj shifts the paradigm from "finding the best way" to "finding all effective ways" to manipulate objects. By providing a diverse library of trajectories, it enables robots to be more robust in open-ended environments. If one trajectory is blocked by an obstacle or a grasp point is slippery, the robot can instantly switch to an alternative primitive from its archive. This work provides a foundational layer for higher-level autonomous task planners to execute complex household chores.