MoDeSuite: Robot Learning Task Suite for Benchmarking Mobile Manipulation with Deformable Objects

Imagine a world where robots aren't just clumsy boxes that can only push heavy, unyielding crates. Imagine them as dexterous helpers capable of tying shoelaces, folding laundry, or navigating a room full of hanging curtains. This is the frontier of mobile manipulation, a field of robotics that combines two difficult skills: moving around (like a dog or a car) and using arms to grab things (like a human). While robots have gotten pretty good at moving rigid objects like boxes or doors, they still struggle miserably with deformable objects. These are things that squish, stretch, bend, or flop around, like a rubber band, a wet towel, or a piece of rope. Because these objects change shape every time you touch them, they are a nightmare for computer programs trying to predict what will happen next. If a robot can't handle a floppy blanket, it can't truly help us in our messy, real-world homes.

This is where a new project called MoDeSuite steps in. Think of it as a high-tech "video game" designed specifically to train robots to handle floppy things. The researchers created a digital playground filled with eight different challenges, ranging from dragging a stretchy belt over a hurdle to pulling a tablecloth off a table without knocking over the dishes. They tested this playground with two types of robots: one with wheels (like a smart vacuum cleaner with arms) and one with legs (a four-legged robot that looks like a dog). The goal was to see if robots could learn to move their bodies and arms together to solve these squishy puzzles. The results show that while robots can learn to do this in the computer game, moving them to the real world is still a tricky business, especially when it comes to seeing and reacting to visual changes.

The Paper's Story: Teaching Robots to Dance with Floppy Things

The paper introduces MoDeSuite, the first standardized "task suite" (a collection of challenges) designed to benchmark how well robots can move around while handling objects that bend and stretch. Before this, researchers had to build their own custom simulations for every new experiment, which made it hard to compare different robots or learning methods. MoDeSuite fixes this by offering a shared, high-quality digital environment where anyone can test their robot learning algorithms.

The suite features eight distinct tasks that cover two main types of "floppy" materials:

1. Elastic objects: Things that stretch and snap back, like rubber bands or foam noodles.
2. Plastic objects: Things that drape and fold but don't snap back, like curtains, tablecloths, or ropes.

The tasks are designed to be tricky. For example, in the "Bend" task, a robot must carry a long, stretchy rod through an L-shaped hallway. If the robot moves too fast or holds the rod wrong, the rod might hit the walls. In the "Drag" task, the robot has to lift a stretchy belt over an obstacle and place it on the other side, all while keeping its own body balanced. In the "Curtain" task, the robot must pull a hanging curtain aside and walk through the opening without bumping into it. These tasks force the robot to coordinate its base (wheels or legs) with its arm, using the object's flexibility to its advantage rather than fighting against it.

To test if this system works, the authors trained four different "learning brains" (algorithms) inside the simulation. Two of these were Reinforcement Learning (RL) algorithms, which learn by trial and error, trying to get a high score. The other two were Imitation Learning (IL) algorithms, which learn by watching a human (or a keyboard controller) perform the task and trying to copy the moves.

What they found in the simulation:
The robots were surprisingly successful in the digital world. The PPO algorithm (a type of RL) generally performed better than the SAC algorithm across most tasks. However, the type of robot mattered a lot. The wheeled robot (Franka on a Ridgeback base) found the tasks easier, while the legged robot (Boston Dynamics' Spot) struggled more, often losing its balance or colliding with walls in tight spaces. This suggests that balancing a robot while it manipulates a floppy object is a massive challenge, especially for legged machines.

The Real-World Test (Sim-to-Real):
The most exciting part of the paper is what happened when they took the trained robots out of the computer and onto the real floor. They deployed the policies onto a real Spot robot without any extra fine-tuning.

• Success with "State" data: When the robot used direct sensor data (knowing exactly where the object's points were in 3D space), the transfer was impressive. For the "Place" and "Drag" tasks, the real robot achieved success rates of 90% to 100%, very close to its performance in the simulation. It could successfully drag a rubber belt over a hurdle in the real world.
• Failure with "Image" data: When the robot tried to learn using only camera images (like a human seeing the world), it struggled. For the "Curtain" task, the robot failed completely in the real world, even though it did well in the simulation. The researchers found that the robot's "eyes" saw the simulation and the real world as completely different things. The visual gap was too wide for the robot to bridge without extra training.

What this means:
The paper doesn't claim to have solved the problem of mobile manipulation with deformable objects. Instead, it provides a crucial new tool: a standardized, challenging benchmark that reveals exactly where current robots succeed and where they fail. It proves that robots can learn to manipulate floppy objects in simulation and that this knowledge can transfer to the real world, but only if the robot has precise information about the object's shape. If the robot relies only on what it sees with a camera, the "uncanny valley" between the simulation and reality is still too deep to cross.

The authors suggest that future work needs to focus on helping robots understand visual data better and on making legged robots more stable when handling heavy or floppy loads. By opening up this "MoDeSuite" to the public, they hope to accelerate research, turning the dream of a robot that can fold your laundry or untangle your headphones into a reality.