Embodiment-conditioned Generalist Control for Multirotor Aerial Robots

Problem
This work addresses the lack of a unified control policy capable of managing various multirotor configurations, such as quadrotors and hexarotors, with a single set of weights. Existing methods typically require separate training for each configuration, leading to inefficiencies and limitations in adaptability. The authors present a preprint, indicating that the work has not yet undergone peer review.

Method
The core contribution is a generalist position control policy that utilizes an embodiment descriptor, specifically a mass and inertia-normalized control allocation matrix. This matrix effectively captures the relationship between motor thrusts and the resulting linear and angular accelerations in the body frame of the multirotor. The policy is trained using Proximal Policy Optimization (PPO) on a diverse set of multirotor configurations, including non-planar and asymmetric designs. The training process is efficient, requiring only five minutes on an NVIDIA RTX 3090 GPU, facilitated by a custom NVIDIA Warp-based dynamics simulator. The architecture of the neural network is compact, allowing for rapid adaptation to various configurations without the need for retraining.

Results
The authors demonstrate that the embodiment-conditioned policy achieves robust control across multiple morphologies in extensive simulation experiments. Notably, the policy exhibits zero-shot transfer capabilities, successfully controlling three distinct hexarotor systems in real-world scenarios: a planar robot, a partially symmetric non-planar system, and a random asymmetric non-planar configuration. The results indicate that the policy maintains high performance across these diverse configurations, showcasing its generalist capabilities.

Limitations
The authors acknowledge that the approach may not generalize to configurations outside the sampled distribution, particularly those with extreme physical characteristics not represented in the training set. Additionally, the reliance on simulation for training may introduce discrepancies when transitioning to real-world applications, although the zero-shot transfer results are promising. The paper does not address potential limitations related to the scalability of the method to larger or more complex multirotor systems.

Why it matters
This research has significant implications for the field of robotics, particularly in the development of adaptable control systems for aerial vehicles. The ability to control various multirotor configurations with a single policy can streamline deployment in diverse environments and applications, reducing the need for extensive retraining. This work contributes to the ongoing discourse on generalist policies in reinforcement learning and robotics, as highlighted in related literature. The findings are available on arXiv, providing a foundation for future research in adaptable control systems for aerial robotics.