Yunlong Song

Yunlong Song, Angel Romero, Matthias Mueller et al.

A central question in robotics is how to design a control system for an agile mobile robot. This paper studies this question systematically, focusing on a challenging setting: autonomous drone racing. We show that a neural network controller trained with reinforcement learning (RL) outperformed optimal control (OC) methods in this setting. We then investigated which fundamental factors have contributed to the success of RL or have limited OC. Our study indicates that the fundamental advantage of RL over OC is not that it optimizes its objective better but that it optimizes a better objective. OC decomposes the problem into planning and control with an explicit intermediate representation, such as a trajectory, that serves as an interface. This decomposition limits the range of behaviors that can be expressed by the controller, leading to inferior control performance when facing unmodeled effects. In contrast, RL can directly optimize a task-level objective and can leverage domain randomization to cope with model uncertainty, allowing the discovery of more robust control responses. Our findings allowed us to push an agile drone to its maximum performance, achieving a peak acceleration greater than 12 times the gravitational acceleration and a peak velocity of 108 kilometers per hour. Our policy achieved superhuman control within minutes of training on a standard workstation. This work presents a milestone in agile robotics and sheds light on the role of RL and OC in robot control.

Yunlong Song, Kexin Shi, Robert Penicka et al.

Recently, neural control policies have outperformed existing model-based planning-and-control methods for autonomously navigating quadrotors through cluttered environments in minimum time. However, they are not perception aware, a crucial requirement in vision-based navigation due to the camera's limited field of view and the underactuated nature of a quadrotor. We propose a learning-based system that achieves perception-aware, agile flight in cluttered environments. Our method combines imitation learning with reinforcement learning (RL) by leveraging a privileged learning-by-cheating framework. Using RL, we first train a perception-aware teacher policy with full-state information to fly in minimum time through cluttered environments. Then, we use imitation learning to distill its knowledge into a vision-based student policy that only perceives the environment via a camera. Our approach tightly couples perception and control, showing a significant advantage in computation speed (10 times faster) and success rate. We demonstrate the closed-loop control performance using hardware-in-the-loop simulation.

Jiawei Fu, Yunlong Song, Yan Wu et al.

Autonomous drones can operate in remote and unstructured environments, enabling various real-world applications. However, the lack of effective vision-based algorithms has been a stumbling block to achieving this goal. Existing systems often require hand-engineered components for state estimation, planning, and control. Such a sequential design involves laborious tuning, human heuristics, and compounding delays and errors. This paper tackles the vision-based autonomous-drone-racing problem by learning deep sensorimotor policies. We use contrastive learning to extract robust feature representations from the input images and leverage a two-stage learning-by-cheating framework for training a neural network policy. The resulting policy directly infers control commands with feature representations learned from raw images, forgoing the need for globally-consistent state estimation, trajectory planning, and handcrafted control design. Our experimental results indicate that our vision-based policy can achieve the same level of racing performance as the state-based policy while being robust against different visual disturbances and distractors. We believe this work serves as a stepping-stone toward developing intelligent vision-based autonomous systems that control the drone purely from image inputs, like human pilots.

Jiaxu Xing, Leonard Bauersfeld, Yunlong Song et al. · eth-zurich

Scene transfer for vision-based mobile robotics applications is a highly relevant and challenging problem. The utility of a robot greatly depends on its ability to perform a task in the real world, outside of a well-controlled lab environment. Existing scene transfer end-to-end policy learning approaches often suffer from poor sample efficiency or limited generalization capabilities, making them unsuitable for mobile robotics applications. This work proposes an adaptive multi-pair contrastive learning strategy for visual representation learning that enables zero-shot scene transfer and real-world deployment. Control policies relying on the embedding are able to operate in unseen environments without the need for finetuning in the deployment environment. We demonstrate the performance of our approach on the task of agile, vision-based quadrotor flight. Extensive simulation and real-world experiments demonstrate that our approach successfully generalizes beyond the training domain and outperforms all baselines.

Nico Messikommer, Yunlong Song, Davide Scaramuzza

In Reinforcement Learning, the trade-off between exploration and exploitation poses a complex challenge for achieving efficient learning from limited samples. While recent works have been effective in leveraging past experiences for policy updates, they often overlook the potential of reusing past experiences for data collection. Independent of the underlying RL algorithm, we introduce the concept of a Contrastive Initial State Buffer, which strategically selects states from past experiences and uses them to initialize the agent in the environment in order to guide it toward more informative states. We validate our approach on two complex robotic tasks without relying on any prior information about the environment: (i) locomotion of a quadruped robot traversing challenging terrains and (ii) a quadcopter drone racing through a track. The experimental results show that our initial state buffer achieves higher task performance than the nominal baseline while also speeding up training convergence.

Yunlong Song, Sangbae Kim, Davide Scaramuzza

This work explores the potential of using differentiable simulation for learning quadruped locomotion. Differentiable simulation promises fast convergence and stable training by computing low-variance first-order gradients using robot dynamics. However, its usage for legged robots is still limited to simulation. The main challenge lies in the complex optimization landscape of robotic tasks due to discontinuous dynamics. This work proposes a new differentiable simulation framework to overcome these challenges. Our approach combines a high-fidelity, non-differentiable simulator for forward dynamics with a simplified surrogate model for gradient backpropagation. This approach maintains simulation accuracy by aligning the robot states from the surrogate model with those of the precise, non-differentiable simulator. Our framework enables learning quadruped walking in simulation in minutes without parallelization. When augmented with GPU parallelization, our approach allows the quadruped robot to master diverse locomotion skills on challenging terrains in minutes. We demonstrate that differentiable simulation outperforms a reinforcement learning algorithm (PPO) by achieving significantly better sample efficiency while maintaining its effectiveness in handling large-scale environments. Our method represents one of the first successful applications of differentiable simulation to real-world quadruped locomotion, offering a compelling alternative to traditional RL methods.

Jiaxu Xing, Ismail Geles, Yunlong Song et al.

Reinforcement learning (RL) has shown great effectiveness in quadrotor control, enabling specialized policies to develop even human-champion-level performance in single-task scenarios. However, these specialized policies often struggle with novel tasks, requiring a complete retraining of the policy from scratch. To address this limitation, this paper presents a novel multi-task reinforcement learning (MTRL) framework tailored for quadrotor control, leveraging the shared physical dynamics of the platform to enhance sample efficiency and task performance. By employing a multi-critic architecture and shared task encoders, our framework facilitates knowledge transfer across tasks, enabling a single policy to execute diverse maneuvers, including high-speed stabilization, velocity tracking, and autonomous racing. Our experimental results, validated both in simulation and real-world scenarios, demonstrate that our framework outperforms baseline approaches in terms of sample efficiency and overall task performance.

Yunlong Song, Davide Scaramuzza

Policy Search and Model Predictive Control~(MPC) are two different paradigms for robot control: policy search has the strength of automatically learning complex policies using experienced data, while MPC can offer optimal control performance using models and trajectory optimization. An open research question is how to leverage and combine the advantages of both approaches. In this work, we provide an answer by using policy search for automatically choosing high-level decision variables for MPC, which leads to a novel policy-search-for-model-predictive-control framework. Specifically, we formulate the MPC as a parameterized controller, where the hard-to-optimize decision variables are represented as high-level policies. Such a formulation allows optimizing policies in a self-supervised fashion. We validate this framework by focusing on a challenging problem in agile drone flight: flying a quadrotor through fast-moving gates. Experiments show that our controller achieves robust and real-time control performance in both simulation and the real world. The proposed framework offers a new perspective for merging learning and control.

Yunlong Song, HaoChih Lin, Elia Kaufmann et al.

Professional race-car drivers can execute extreme overtaking maneuvers. However, existing algorithms for autonomous overtaking either rely on simplified assumptions about the vehicle dynamics or try to solve expensive trajectory-optimization problems online. When the vehicle approaches its physical limits, existing model-based controllers struggle to handle highly nonlinear dynamics, and cannot leverage the large volume of data generated by simulation or real-world driving. To circumvent these limitations, we propose a new learning-based method to tackle the autonomous overtaking problem. We evaluate our approach in the popular car racing game Gran Turismo Sport, which is known for its detailed modeling of various cars and tracks. By leveraging curriculum learning, our approach leads to faster convergence as well as increased performance compared to vanilla reinforcement learning. As a result, the trained controller outperforms the built-in model-based game AI and achieves comparable overtaking performance with an experienced human driver.

Yunlong Song, Mats Steinweg, Elia Kaufmann et al.

In many robotic tasks, such as autonomous drone racing, the goal is to travel through a set of waypoints as fast as possible. A key challenge for this task is planning the time-optimal trajectory, which is typically solved by assuming perfect knowledge of the waypoints to pass in advance. The resulting solution is either highly specialized for a single-track layout, or suboptimal due to simplifying assumptions about the platform dynamics. In this work, a new approach to near-time-optimal trajectory generation for quadrotors is presented. Leveraging deep reinforcement learning and relative gate observations, our approach can compute near-time-optimal trajectories and adapt the trajectory to environment changes. Our method exhibits computational advantages over approaches based on trajectory optimization for non-trivial track configurations. The proposed approach is evaluated on a set of race tracks in simulation and the real world, achieving speeds of up to 60 km/h with a physical quadrotor.

Yunlong Song, Selim Naji, Elia Kaufmann et al.

State-of-the-art quadrotor simulators have a rigid and highly-specialized structure: either are they really fast, physically accurate, or photo-realistic. In this work, we propose a novel quadrotor simulator: Flightmare. Flightmare is composed of two main components: a configurable rendering engine built on Unity and a flexible physics engine for dynamics simulation. Those two components are totally decoupled and can run independently of each other. This makes our simulator extremely fast: rendering achieves speeds of up to 230 Hz, while physics simulation of up to 200,000 Hz on a laptop. In addition, Flightmare comes with several desirable features: (i) a large multi-modal sensor suite, including an interface to extract the 3D point-cloud of the scene; (ii) an API for reinforcement learning which can simulate hundreds of quadrotors in parallel; and (iii) integration with a virtual-reality headset for interaction with the simulated environment. We demonstrate the flexibility of Flightmare by using it for two different robotic tasks: quadrotor control using deep reinforcement learning and collision-free path planning in a complex 3D environment.

Florian Fuchs, Yunlong Song, Elia Kaufmann et al.

Autonomous car racing is a major challenge in robotics. It raises fundamental problems for classical approaches such as planning minimum-time trajectories under uncertain dynamics and controlling the car at the limits of its handling. Besides, the requirement of minimizing the lap time, which is a sparse objective, and the difficulty of collecting training data from human experts have also hindered researchers from directly applying learning-based approaches to solve the problem. In the present work, we propose a learning-based system for autonomous car racing by leveraging a high-fidelity physical car simulation, a course-progress proxy reward, and deep reinforcement learning. We deploy our system in Gran Turismo Sport, a world-leading car simulator known for its realistic physics simulation of different race cars and tracks, which is even used to recruit human race car drivers. Our trained policy achieves autonomous racing performance that goes beyond what had been achieved so far by the built-in AI, and, at the same time, outperforms the fastest driver in a dataset of over 50,000 human players.

Yunlong Song, Davide Scaramuzza

The combination of policy search and deep neural networks holds the promise of automating a variety of decision-making tasks. Model Predictive Control (MPC) provides robust solutions to robot control tasks by making use of a dynamical model of the system and solving an optimization problem online over a short planning horizon. In this work, we leverage probabilistic decision-making approaches and the generalization capability of artificial neural networks to the powerful online optimization by learning a deep high-level policy for the MPC (High-MPC). Conditioning on robot's local observations, the trained neural network policy is capable of adaptively selecting high-level decision variables for the low-level MPC controller, which then generates optimal control commands for the robot. First, we formulate the search of high-level decision variables for MPC as a policy search problem, specifically, a probabilistic inference problem. The problem can be solved in a closed-form solution. Second, we propose a self-supervised learning algorithm for learning a neural network high-level policy, which is useful for online hyperparameter adaptations in highly dynamic environments. We demonstrate the importance of incorporating the online adaption into autonomous robots by using the proposed method to solve a challenging control problem, where the task is to control a simulated quadrotor to fly through a swinging gate. We show that our approach can handle situations that are difficult for standard MPC.