PhyRoGen Generates 24 Robot Manipulation Puzzles in Seconds

Manipulation puzzles—where one object must be moved before another—are crucial for training robots in assembly and disassembly. But generating large, diverse datasets for these tasks is typically manual and time-consuming. To address this, researchers from TU Berlin (Lennart Julian Droß, Andreas Orthey, Marc Toussaint) developed PhyRoGen, a framework that uses procedural content generation (PCG) to automatically produce synthetic datasets of physical manipulation puzzles. PCG, commonly used in video games, generates content algorithmically rather than by hand. PhyRoGen creates puzzles with interlocking object dependencies, meaning an articulated object must be manipulated before another can be moved. The framework defines six concrete puzzle generators, which together produce 24 distinct physical puzzles. Using sampling-based planning algorithms, all puzzles are solvable within 1 to 300 seconds.

The researchers validated their approach by demonstrating that every generated puzzle can be physically manipulated by a KUKA LBR iiwa robot in simulation. This confirms that PhyRoGen can procedurally generate unique, solvable puzzles—a critical ingredient for benchmarking manipulation algorithms and developing robust foundation models. By automating dataset creation, PhyRoGen eliminates the bottleneck of manual puzzle design, enabling scalable and diverse training data. The paper, accepted at CASE 2026, represents a step toward more capable and generalizable robot manipulation systems. For AI and robotics professionals, this means faster iteration on manipulation algorithms and the potential for foundation models trained on procedurally generated tasks—without labor-intensive data collection.