Â鶹ÒùÔº

Astrobee learns to transport soft cargo: Open-source simulator models real ISS challenges

Astrobee is a free-flying robotic system developed by NASA that is made up of three distinct cube-shaped robots. This system was originally designed to help astronauts who are working at the International Space Station (ISS) by automating some of their routine manual tasks.

While Astrobee could be highly valuable for astronauts, boosting the efficiency with which they complete day-to-day operations, its object manipulation capabilities are not yet optimal. Specifically, past experiments suggest that the robot struggles when handling deformable items, including cargo bags that resemble some of those that it might be tasked to pick up on the ISS.

Researchers at Stanford University, University of Cambridge and NASA Ames recently developed Pyastrobee, a simulation environment and control stack to train Astrobee in Python, with a particular emphasis on the manipulation and transport of cargo.

This new simulation and control toolkit, presented in a posted on the arXiv preprint server, was used to train Astrobee to successfully transfer cargo between different ISS modules, without colliding with other objects.

"We are lucky to collaborate with NASA Ames, and one of the problems they're interested in is how to have Astrobee (their free-floating space robot in the International Space Station) perform logistics and maintenance tasks," Daniel Morton, first author of the paper, told Â鶹ÒùÔº.

"This is especially important for any future space stations that may not be continuously crewed, and will require autonomous operations from these robots for 'chores' like re-stocking the station with cargo. However, getting Astrobee to manipulate and move around these cargo bags is a really hard problem to solve."

A key challenge that has so far limited the effectiveness of Astrobee in carrying and manipulating cargo bags is that these bags are typically made of a deformable vinyl-based material. Predicting exactly how the cargo bags will deform when the robot grasps them and interact with them can be very difficult.

"We set out to find a way to control this and built a simulation environment that can not only accurately represent the ISS, but also model deformable cargo," explained Morton. "Pyastrobee is unique due to its modeling of deformable cargo, since we use a physics engine (Bullet) that allows this. It is also developed in Python, making it easy to rapidly prototype different controllers, and integrate with other robotics software tools."

Morton and his colleagues integrated their simulation environment with reinforcement learning (RL) software known as Gymnasium and Stable Baselines. Their hope was that this software would facilitate the use of their platform for testing RL-based object manipulation strategies in space.

"We've found that a simulator-in-the-loop sampling-based model-predictive-controller (MPC) is a good preliminary approach for this problem," said Morton. "Using the simulator as the model makes it easier to specify how Astrobee and the cargo bag move together, rather than trying to derive a challenging closed-form model of the system. We've also experimented with models of different fidelity, exploring the trade-offs between computational accuracy and speed."

Pyastrobee, the simulator developed by this research team, could soon be used both by engineers and students to test their space robotics algorithms. The code for the simulator, as well as its integrated control and planning methods, is open-source and can be .

"Now that we have a preliminary approach (sampling-based MPC), I'd like to explore how to make this much more computationally efficient," added Morton. "I've recently been working on and this would be perfect to put on Astrobee, to give guarantees on constraints like collision avoidance. I'd also like to explore how to use multiple Astrobees to perform this task—two Astrobees holding either side of the bag will likely improve stability during transport."