Earlier this week, Alphabet-owned DeepMind acquired a physics simulation platform MuJoCo, which stands for Multi-Joint Dynamics with Contact.
After the acquisition, the DeepMind Robotics Simulation team, which had been using MuJoCo in the past, is planning to fully open-source the platform in 2022 and make it freely available for everyone to support research everywhere.
Sign up for your weekly dose of what's up in emerging technology.
MuJoCo was first developed by Emo Todorov for Roboti and was available as a commercial product from 2015 to 2021. After DeepMind acquired MuJoCo, it is making it freely available to everyone. However, the details of the financial transactions are yet to be disclosed.
Post-acquisition, Roboti will continue to support existing paid licenses until they expire. In addition, the legacy MuJoCo release (versions 2.0 and earlier) will remain available for download, with a free activation key file, valid until October 2031.
What is MuJoCo?
MuJoCo is a physics engine that aims to facilitate research and development in robotics, graphics, biomechanics, animation, and other domains requiring fast and accurate simulation. It is one of the first full-featured simulators designed from scratch for model-based optimisation, particularly through contacts.
The platform makes it possible to scale up computationally intensive techniques such as optimal control, physically consistent state estimation, system identification and automated mechanism design and apply them to complex dynamical systems in contact-rich behaviours. Plus, it has more traditional applications such as testing and validating control schemes before deployment on physical robots, interactive scientific visualisation, virtual environments, animation, and gaming.
How is MuJoCo different?
DeepMind MuJoCo is not alone. Other simulator platforms include Facebook’s Habitat 2.0 and AI2’s ManipulaTHOR. However, what sets them apart is its contact model, which accurately and efficiently captures the salient features of contacting objects. Like other rigid-body simulators, it avoids the fine details of deformations at the contact site and often runs much faster than in real-time.
“Unlike other simulators, MuJoCo resolves contact forces using the convex Gauss Principle,” said the DeepMind Robotics Simulation team. The convexity ensures unique solutions and well-defined inverse dynamics. Plus, the model is flexible, providing multiple parameters which are tuned to approximate a wide range of contact phenomena.
Further, the DeepMind team said that their platform is based on real physics and takes no shortcuts. According to them, many simulations were initially designed for purposes like gaming and cinema; they sometimes take shortcuts that prioritise stability over accuracy. For example, they may ignore gyroscopic forces or directly modify velocities.
That, in the context of optimisation, can be particularly harmful. In contrast, MuJoCo is a second-order continuous-time simulator, implementing the full equations of motion. In other words, MuJoCo closely adheres to the equations that govern our world — non-trivial physical phenomena like Newton’s Cradle, and unintuitive ones like the Dzhanibekov effect, happens naturally.
The team also said that the MuJoCo core engine is written in pure C, making it easily portable to various architectures. In addition to this, the platform also provides fast and convenient computations of commonly used quantities, like kinematic Jacobians and inertia matrices.
MuJoCo offers powerful scene descriptions. It uses cascading defaults – avoiding multiple repeated values – and contains elements for real-world robotic components like tendons, actuators, equality constraints, motion-capture markers, and sensors. Soon, it plans to include standardising MJCF as an open format to extend its usefulness beyond the MuJoCo ecosystem.
Besides this, MuJoCo includes two powerful features that support musculoskeletal models of humans and animals. It captures the complexity of biological muscles, including activation states and force-length-velocity curves.
DeepMind Slaying Robotics
The recent acquisition comes at a time when there is a dearth of data in robotics research. This is one of the reasons why DeepMind’s arch-rival OpenAI went on to shut down its robotics arm indefinitely. But, this is not stopping DeepMind, as its teams are trying to get around this paucity of data with a technique called sim-to-real, in a big way.
Now, with the acquisition of MuJoCo, open-sourcing the library seems like a smooth move for the company, and surely going to benefit the robotics ecosystem as a whole.