Atil Iscen
Research Areas
Authored Publications
Sort By
Robotic Table Tennis: A Case Study into a High Speed Learning System
Jon Abelian
Saminda Abeyruwan
Michael Ahn
Justin Boyd
Erwin Johan Coumans
Omar Escareno
Wenbo Gao
Navdeep Jaitly
Juhana Kangaspunta
Satoshi Kataoka
Gus Kouretas
Yuheng Kuang
Corey Lynch
Thinh Nguyen
Ken Oslund
Barney J. Reed
Anish Shankar
Avi Singh
Grace Vesom
Peng Xu
Robotics: Science and Systems (2023)
Preview abstract
We present a deep-dive into a learning robotic system that, in previous work, was shown to be capable of hundreds of table tennis rallies with a human and has the ability to precisely return the ball to desired targets. This system puts together a highly optimized and novel perception subsystem, a high-speed low-latency robot controller, a simulation paradigm that can prevent damage in the real world and also train policies for zero-shot transfer, and automated real world environment resets that enable autonomous training and evaluation on physical robots. We complement a complete system description including numerous design decisions that are typically not widely disseminated, with a collection of ablation studies that clarify the importance of mitigating various sources of latency, accounting for training and deployment distribution shifts, robustness of the perception system, and sensitivity to policy hyper-parameters and choice of action space. A video demonstrating the components of our system and details of experimental results is included in the supplementary material.
View details
Adversarial Motion Priors Make Good Substitutes for Complex Reward Functions
Ale Escontrela
Jason Peng
Ken Goldberg
Pieter Abbeel
2022 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS (2022) (to appear)
Preview abstract
Training high-dimensional simulated agents with under-specified reward functions often leads to jerky and unnatural behaviors, which results in physically infeasible strategies that are generally ineffective when deployed in the real world. To mitigate these unnatural behaviors, reinforcement learning (RL) practitioners often utilize complex reward functions that encourage more physically plausible behaviors, in conjunction with tricks such as domain randomization to train policies that satisfy the user's style criteria and can be successfully deployed on real robots. Such an approach has been successful in the realm of legged locomotion, leading to state-of-the-art results. However, designing effective reward functions can be a labour-intensive and tedious tuning process, and these hand-designed rewards do not easily generalize across platforms and tasks. We propose substituting complex reward functions with "style rewards" learned from a dataset of motion capture demonstrations. This learned style reward can be combined with a simple task reward to train policies that perform tasks using naturalistic strategies. These more natural strategies can also facilitate transfer to the real world. We build upon prior work in computer graphics and demonstrate that an adversarial approach to training control policies can produce behaviors that transfer to a real quadrupedal robot without requiring complex reward functions. We also demonstrate that an effective style reward can be learned from a few seconds of motion capture data gathered from a German Shepherd and leads to energy-efficient locomotion strategies with natural gait transitions.
View details
Style-Augmented Mutual Information for Practical Skill Discovery
Ale Escontrela
Jason Peng
Ken Goldberg
Pieter Abbeel
Proceedings of NeurIPS (2022) (to appear)
Preview abstract
Exploration and skill discovery in many real-world settings is often inspired by the activities we see others perform. However, most unsupervised skill discovery methods tend to focus solely on the intrinsic component of motivation, often by maximizing the Mutual Information (MI) between the agent's skills and the observed trajectories. These skills, though diverse in the behaviors they elicit, leave much to be desired. Namely, skills learned by maximizing MI in a high-dimensional continuous control setting tend to be aesthetically unpleasing and challenging to utilize in a practical setting, as the violent behavior often exhibited by these skills would not transfer well to the real world. We argue that solely maximizing MI is insufficient if we wish to discover useful skills, and that a notion of "style" must be incorporated into the objective. To this end, we propose the Style-Augmented Mutual Information objective (SAMI), whereby - in addition to maximizing a lower-bound of the MI - the agent is encouraged to minimize the f-divergence between the policy-induced trajectory distribution and the trajectory distribution contained in the reference data (the style objective). We compare SAMI to other popular skill discovery objectives, and demonstrate that skill-conditioned policies optimized with SAMI achieve equal or greater performance when applied to downstream tasks. We also show that the data-driven motion prior specified by the style objective can be inferred from various modalities, including large motion capture datasets or even RGB videos.
View details
Reward Machines for Vision-Based Robotic Manipulation.
Alberto Camacho
Andy Zeng
Dmitry Kalashnikov
Jake Varley
International Conference on Robotics and Automation (2021)
Preview abstract
Deep Q learning (DQN) has enabled robot agents to accomplish vision based tasks that seemed out of reach. Despite recent success stories, there are still several sources of computational complexity that challenge the performance of DQN. We place the focus on vision manipulation tasks, where the correct action selection is often predicated on a small number of pixels. We observe that in some of these tasks DQN does not converge to the optimal Q function, and their values do not separate well optimal and suboptimal actions. In consequence, the policies obtained with DQN tend to be brittle and manifest a low success rate, especially in long horizon tasks. In this work we show the benefits of Reward Machines (RMs) for Deep Q learning (DQRM) in vision based robot manipulation tasks. Reward machines decompose the task at an abstract level, inform the agent about their current stage along task completion, and guide them via dense rewards. We show that RMs help DQN learn the optimal Q values in each abstract state. Their policies are more robust, manifest higher success rate, and are learned with fewer training steps compared with DQN. The benefits of RMs are more evident in long-horizon tasks, where we show that DQRM is able to learn good-quality policies with six times times fewer training steps than DQN, even when this is equipped with dense reward shaping.
View details
Learning to walk on complex terrains with vision
Ale Escontrela
Erwin Johan Coumans
Peng Xu
Sehoon Ha
Conference on Robotic Learning (2021)
Preview abstract
Visual feedback is crucial for legged robots to safely and efficiently handle uneven terrains such as stairs. However, effectively training robots to effectively consume high dimensional visual input for locomotion is challenging. In this work, we propose a framework to train a vision-based locomotion controller for quadruped robots to traverse a variety of uneven environments. Our key idea is to model the locomotion controller as a hierarchical structure with a high-level vision policy and a low-level motion controller. The high-level vision policy takes as input the perceived vision inputs as well as robot states and outputs desired foothold placement and base movement of the robot, which is realized by low level motion controller composed of a position controller for swing legs and a MPC-based torque controller for stance legs. We train the vision policy using Deep Reinforcement Learning and demonstrate our approach on a variety of uneven environments such as step-stones, stairs, pillars, and moving platforms. We also deploy our policy on a real quadruped robot to walk over a series of random step-stones.
View details
Disentangled Planning and Control in Vision Based Robotics via Reward Machines
Alberto Camacho
Andy Zeng
Dmitry Kalashnikov
Jake Varley
Deep Reinforcement Learning Workshop (Deep RL), collocated with NeurIPS 2020 (2020)
Preview abstract
In this work we augment a Deep Q-Learning agent with a Reward Machine (DQRM) to increase speed of learning vision-based policies for robot tasks, and overcome some of the limitations of DQN that prevent it from converging to good-quality policies. A reward machine (RM) is a finite state machine that decomposes a task into a discrete planning graph and equips the agent with a reward function to guide it toward task completion. The reward machine can be used for both reward shaping, and informing the policy what abstract state it is currently at. An abstract state is a high level simplification of the current state, defined in terms of task relevant features. These two supervisory signals of reward shaping and knowledge of current abstract state coming from the reward machine complement each other and can both be used to improve policy performance as demonstrated on several vision based robotic pick and place tasks. Particularly for vision based robotics applications, it is often easier to build a reward machine than to try and get a policy to learn the task without this structure.
View details
Provably Robust Blackbox Optimization for Reinforcement Learning
Aldo Pacchiano
Jack Parker-Holder
Yunhao Tang
Yuxiang Yang
Conference on Robot Learning (CoRL 2019)
Preview abstract
Interest in derivative-free optimization (DFO) and "evolutionary strategies" (ES) has recently surged in the Reinforcement Learning (RL) community, with growing evidence that they can match state of the art methods for policy optimization problems in Robotics. However, it is well known that DFO methods suffer from prohibitively high sampling complexity. They can also be very sensitive to noisy rewards and stochastic dynamics. In this paper, we propose a new class of algorithms, called Robust Blackbox Optimization (RBO). Remarkably, even if up to 23% of all the measurements are arbitrarily corrupted, RBO can provably recover gradients to high accuracy. RBO relies on learning gradient flows using robust regression methods to enable off-policy updates. On several MuJoCo robot control tasks, when all other RL approaches collapse in the presence of adversarial noise, RBO is able to train policies effectively. We also show that RBO can be applied to legged locomotion tasks including path tracking for quadruped robots.
View details
Sim-to-Real: Learning Agile Locomotion For Quadruped Robots
Erwin Coumans
Danijar Hafner
Steven Bohez
RSS (2018)
Preview abstract
Designing agile locomotion for quadruped robots often requires extensive expertise and tedious manual tuning. In this paper, we present a system to automate this process by leveraging deep reinforcement learning techniques. Our system can learn quadruped locomotion from scratch with simple reward signals. In addition, users can provide an open loop reference to guide the learning process if more control over the learned gait is needed. The control policies are learned in a physical simulator and then deployed to real robots. In robotics, policies trained in simulation often does not transfer to the real world. We narrow this reality gap by improving the physical simulator and learning robust policies. We improve the simulation using system identification, developing an accurate actuator model and simulating latency. We learn robust controllers by randomizing the physical environments, adding perturbations and designing a compact observation space. We evaluate our system on two agile locomotion gaits: trotting and galloping. After learning in simulation, a quadruped robot can successfully perform both gaits in real world.
View details
Preview abstract
We propose a simple drop-in noise-tolerant replacement for the standard finite difference procedure used ubiquitously in blackbox optimization. In our approach, parameter perturbation directions are defined by a family of deterministic or randomized structured matrices. We show that at the small cost of computing a Fast Fourier Transform (FFT), such structured finite differences consistently give higher quality approximation of gradients and Jacobians in comparison to vanilla approaches that use coordinate directions or random Gaussian perturbations. We show that
linearization of noisy, blackbox dynamics using our methods leads to improved performance of trajectory optimizers like Iterative LQR and Differential Dynamic Programming on several classic continuous control tasks. By embedding structured exploration in implicit filtering methods, we are able to learn agile walking and turning policies for quadruped locomotion, that successfully transfer from simulation to actual hardware.
We give a theoretical justification of our methods in terms of bounds on the quality of gradient reconstruction in the presence of noise.
View details
Policies Modulating Trajectory Generators
Erwin Coumans
2nd Annual Conference on Robot Learning, CoRL 2018, PMLR, pp. 916-926
Preview abstract
We propose an architecture for learning complex controllable behaviors by having simple Policies Modulate Trajectory Generators (PMTG), a powerful combination that can provide both memory and prior knowledge to the controller. The result is a flexible architecture that is applicable to a class of problems with periodic motion for which one has an insight into the class of trajectories that might lead to a desired behavior. We illustrate the basics of our architecture using a synthetic control problem, then go on to learn speed-controlled locomotion for a quadrupedal robot by using Deep Reinforcement Learning and Evolutionary Strategies. We demonstrate that a simple linear policy, when paired with a parametric Trajectory Generator for quadrupedal gaits, can induce walking behaviors with controllable speed from 4-dimensional IMU observations alone, and can be learned in under 1000 rollouts. We also transfer these policies to a real robot and show locomotion with controllable forward velocity.
View details
