Safe Reinforcement Learning for Legged Locomotion

Designing control policies for legged locomotion is complex due to the under-actuated and non-continuous robot dynamics. Model-free reinforcement learning provides promising tools to tackle this challenge. However, a major bottleneck of applying model-free reinforcement learning in real world is safety. In this paper, we propose a safe reinforcement learning framework that switches between a safe recovery policy that prevents the robot from entering unsafe states, and a learner policy that is optimized to complete the task. The safe recovery policy takes over the control when the learner policy violates safety constraints, and hands over the control back when there are no future safety violations. We design the safe recovery policy so that it ensures safety of legged locomotion while minimally intervening in the learning process. Furthermore, we theoretically analyze the proposed framework and provide an upper bound on the task performance. We verify the proposed framework in four locomotion tasks on a simulated and real quadrupedal robot: efficient gait, catwalk, two-leg balance, and pacing. On average, our method achieves 48.6% fewer falls and comparable or better rewards than the baseline methods in simulation. When deployed it on real-world quadruped robot, our training pipeline enables 34% improvement in energy efficiency for the efficient gait, 40.9% narrower of the feet placement in the catwalk, and two times more jumping duration in the two-leg balance. Our method achieves less than five falls over the duration of 115 minutes of hardware time.

Legged Robots that Keep on Learning: Fine-Tuning Locomotion Policies in the Real World

Legged robots are physically capable of traversing a wide range of challenging environments, but designing controllers that are sufficiently robust to handle this diversity has been a long-standing challenge in robotics. Reinforcement learning presents an appealing approach for automating the controller design process and has been able to produce remarkably robust controllers when trained in a suitable range of environments. However, it is difficult to predict all likely conditions the robot will encounter during deployment and enumerate them at training-time. What if instead of training controllers that are robust enough to handle any eventuality, we enable the robot to continually learn in any setting it finds itself in? This kind of real-world reinforcement learning poses a number of challenges, including efficiency, safety, and autonomy. To address these challenges, we propose a practical robot reinforcement learning system for fine-tuning locomotion policies in the real world. We demonstrate that a modest amount of real-world training can substantially improve performance during deployment, and this enables a real A1 quadrupedal robot to autonomously fine-tune multiple locomotion skills in a range of environments, including an outdoor lawn and a variety of indoor terrains.

Improving Safety in Deep Reinforcement Learning using Unsupervised Action Planning

One of the key challenges to deep reinforcement learning (deep RL) is to ensure safety at both training and testing phases. In this work, we propose a novel technique of unsupervised action planning to improve the safety of on-policy reinforcement learning algorithms, such as trust region policy optimization (TRPO) or proximal policy optimization (PPO). We design our safety-aware reinforcement learning by storing all the history of "recovery" actions that rescue the agent from dangerous situations into a separate "safety" buffer and finding the best recovery action when the agent encounters similar states. Because this functionality requires the algorithm to query similar states, we implement the proposed safety mechanism using an unsupervised learning algorithm, k-means clustering. We evaluate the proposed algorithm on six robotic control tasks that cover navigation and manipulation. Our results show that the proposed safety RL algorithm can achieve higher rewards compared with multiple baselines in both discrete and continuous control problems. The supplemental video can be found at: https://youtu.be/AFTeWSohILo.

Model-based Motion Imitation for Agile, Diverse and Generalizable Quadupedal Locomotion

Robots operating in human environments need a variety of skills, like slow and fast walking, turning, and side-stepping. However, building robot controllers that can exhibit such a large range of behaviors is challenging, and unsolved. We present an approach that uses a model-based controller for imitating different animal gaits without requiring any real-world fine-tuning. Unlike previous works that learn one policy per motion, we present a unified controller which is capable of generating four different animal gaits on the A1 robot. Our framework includes a trajectory optimization procedure that improves the quality of real-world imitation. We demonstrate our results in simulation and on a real 12-DoF A1 quadruped robot. Our result shows that our approach can mimic four animal motions, and outperform baselines learned per motion.

Solving Challenging Control Problems Using Two-Staged Deep Reinforcement Learning

We present a two-staged deep reinforcement learning algorithm for solving challenging control problems. Deep reinforcement learning (deep RL) has been an effective tool for solving many high-dimensional continuous control problems, but it cannot effectively solve challenging problems with certain properties, such as sparse reward functions or sensitive dynamics. In this work, we propose an approach that decomposes the given problem into two stages: motion planning and motion imitation. The motion planning stage seeks to compute a feasible motion plan with approximated dynamics by directly sampling the state space rather than exploring random control signals. Once the motion plan is obtained, the motion imitation stage learns a control policy that can imitate the given motion plan with realistic sensors and actuations. We demonstrate that our approach can solve challenging control problems - rocket navigation and quadrupedal locomotion - which cannot be solved by the standard MDP formulation. The supplemental video can be found at: https://youtu.be/FYLo1Ov_8-g

Finite State Machine Policies Modulating Trajectory Generator

Deep reinforcement learning (deep RL) has emerged as an effective tool for developing controllers for legged robots. However, a simple neural network representation is known for its poor extrapolation ability, making the learned behavior vulnerable to unseen perturbations or challenging terrains. Therefore, researchers have investigated a novel architecture, Policies Modulating Trajectory Generators (PMTG), which combines trajectory generators (TG) and feedback control signals to achieve more robust behaviors. In this work, we propose to extend the PMTG framework with a finite state machine PMTG by replacing simple TGs with asynchronous finite state machines (Async FSMs). This invention offers an explicit notion of contact events to the policy to negotiate unexpected perturbations. We demonstrated that the proposed architecture could achieve more robust behaviors in various scenarios, such as challenging terrains or external perturbations, on both simulated and real robots. The supplemental video can be found at: http://youtu.be/XUiTSZaM8f0.

Learning Robust Agents for Visual Navigation in Dynamic Environments: The Winning Entry of iGibson Challenge 2021

This paper presents an approach for improving navigation in dynamic and interactive environments, which won the 1st place in the iGibson Interactive Navigation Challenge 2021. While the last few years have produced impressive progress on PointGoal Navigation in static environments, relatively little effort has been made on more realistic dynamic environments. The iGibson Challenge proposed two new navigation tasks, Interactive Navigation and Social Navigation, which add displaceable obstacles and moving pedestrians into the simulator environment. Our approach to study these problems uses two key ideas. First, we employ large-scale reinforcement learning by leveraging the Habitat simulator, which supports high performance parallel computing for both simulation and synchronized learning. Second, we employ a new data augmentation technique that adds more dynamic objects into the environment, which can also be combined with traditional image-based augmentation techniques to boost the performance further. Lastly, we achieve sim-to-sim transfer from Habitat to the iGibson simulator, and demonstrate that our proposed methods allow us to train robust agents in dynamic environments with interactive objects or moving humans. Video link: https://www.youtube.com/watch?v=HxUX2HeOSE4

Graph-based Cluttered Scene Generation and Interactive Exploration using Deep Reinforcement Learning

We introduce a novel method to teach a robotic agent to interactively explore cluttered yet structured scenes, such as kitchen pantries and grocery shelves, by leveraging the physical plausibility of the scene. We propose a novel learning framework to train an effective scene exploration policy to discover hidden objects with minimal interactions. First, we define a novel scene grammar to represent structured clutter. Then we train a Graph Neural Network (GNN) based Scene Generation agent using deep reinforcement learning (deep RL), to manipulate this Scene Grammar to create a diverse set of stable scenes, each containing multiple hidden objects. Given such cluttered scenes, we then train a Scene Exploration agent, using deep RL, to uncover hidden objects by interactively rearranging the scene. We show that our learned agents hide and discover significantly more objects than the baselines. We present quantitative results that prove the generalization capabilities of our agents. We also demonstrate sim-to-real transfer by successfully deploying the learned policy on a real UR10 robot to explore real-world cluttered scenes. The supplemental video can be found at https://www.youtube.com/watch?v=T2Jo7wwaXss.

Learning Robot Structure and Motion Embeddings using Graph Neural Networks

We propose a learning framework to find the representation of a robot's kinematic structure and motion embedding spaces using graph neural networks (GNN). Finding a compact and low-dimensional embedding space for complex phenomena is a key for understanding its behaviors, which may lead to a better learning performance, as we observed in other domains of images or languages. However, although numerous robotics applications deal with various types of data, the embedding of the generated data has been relatively less studied by roboticists. To this end, our work aims to learn embeddings for two types of robotic data: the robot's design structure, such as links, joints, and their relationships, and the motion data, such as kinematic joint positions. Our method exploits the tree structure of the robot to train appropriate embeddings to the given robot data. To avoid overfitting, we formulate multi-task learning to find a general representation of the embedding spaces. We evaluate the proposed learning method on a robot with a simple linear structure and visualize the learned embeddings using t-SNE. We also study a few design choices of the learning framework, such as network architectures and message passing schemes.

Learning to Navigate Sidewalks in Outdoor Environments

Outdoor navigation on sidewalks in urban environments is the key technology behind important human assistive applications, such as last-mile delivery or neighborhood patrol. This paper aims to develop a quadruped robot that follows a route plan generated by public map services, while remaining on sidewalks and avoiding collisions with obstacles and pedestrians. We devise a two-staged learning framework, which first trains a teacher agent in an abstract world with privileged ground-truth information, and then applies Behavior Cloning to teach the skills to a student agent who only has access to realistic sensors. The main research effort of this paper focuses on overcoming challenges when deploying the student policy on a quadruped robot in the real world. We propose methodologies for designing sensing modalities, network architectures, and training procedures to enable zero-shot policy transfer to unstructured and dynamic real outdoor environments. We evaluate our learning framework on a quadrupedal robot navigating sidewalks in the city of Atlanta, USA. Using the learned navigation policy and its onboard sensors, the robot is able to walk 3.2 kilometers with a limited number of human interventions.