We demonstrate the ability of large language models (LLMs) to perform iterative self-improvement of robot policies. An important insight of this paper is that LLMs have a built-in ability to perform (stochastic) numerical optimization and that this property can be leveraged for explainable robot policy search. Based on this insight, we introduce the SAS Prompt (Summarize, Analyze, Synthesize) – a single prompt that enables iterative learning and adaptation of robot behavior by combining the LLM's ability to retrieve, reason and optimize over previous robot traces in order to synthesize new, unseen behavior. Our approach can be regarded as an early example of a new family of explainable policy search methods that are entirely implemented within an LLM. We evaluate our approach both in simulation and on a real-robot table tennis task. Project website: sites.google.com/asu.edu/sas-llm/

Achieving human-level performance on real world tasks is a north star for the robotics community. We present the first learned robot agent that reaches amateur humanlevel performance in competitive table tennis. Table tennis is a physically demanding sport that takes humans years to master. We contribute (1) a hierarchical and modular policy architecture consisting of (i) low level controllers with their skill descriptors that model their capabilities and (ii) a high level controller that chooses the low level skills, (2) techniques for enabling zero-shot sim-to-real and curriculum building, including an iterative approach (train in sim, deploy in real), and (3) real time adaptation to unseen opponents. Policy performance was assessed through 29 robot vs. human matches of which the robot won 45 % (13/29). All humans were unseen players and their skill level varied from beginner to tournament level. Whilst the robot lost all matches vs. the most advanced players it won 100 % matches vs. beginners and 55 % matches vs. intermediate players, demonstrating solidly amateur humanlevel performance. Videos of the matches can be viewed here1.See sites https://google.com/view/competitive-robot-table-tennis.

The use of sparse neural networks has seen rapid growth in recent years, particularly in computer vision. Their appeal stems largely from the reduced number of parameters required to train and store, as well as in an increase in learning efficiency. Somewhat surprisingly, there have been very few efforts exploring their use in Deep Reinforcement Learning (DRL). In this work we perform a systematic investigation into applying a number of existing sparse training techniques on a variety of DRL agents and environments. Our results corroborate the findings from sparse training in the computer vision domain {–}sparse networks perform better than dense networks for the same parameter count{–} in the DRL domain. We provide detailed analyses on how the various components in DRL are affected by the use of sparse networks and conclude by suggesting promising avenues for improving the effectiveness of sparse training methods, as well as for advancing their use in DRL.

Learning goal conditioned control in the real world is a challenging open problem in robotics. Reinforcement learning systems have the potential to learn autonomously via trial-and-error, but in practice the costs of manual reward design, ensuring safe exploration, and hyperparameter tuning are often enough to preclude real world deployment. Imitation learning approaches, on the other hand, offer a simple way to learn control in the real world, but typically require costly cu-rated demonstration data and lack a mechanism for continuous improvement. Recently, iterative imitation methods have been shown to be effective at relaxing both these constraints, learning goal directed control from undirected demonstration data, and improving continuously via self-supervised goal reaching. These approaches, however, have not yet been shown to scale beyond simple simulated environments. In this work, we present the first evidence that simple iterative imitation learning can scale to goal-directed behavior on a real robot in a dynamic setting: high speed, precision table tennis (e.g. “land the ball on this particular target”). We find that this approach offers a straightforward way to do continuous on-robot learning, without complexities such as reward design, value function learning, or sim-to-real transfer. We also find that this approach is scalable-sample efficient enough to train on a physical robot in just a few hours. In real world evaluations, we find that that the resulting policy can perform on par or better than amateur humans (with players sampled randomly from a robotics lab) at the task of returning the ball to specific targets on the table. Finally, we analyze the effect of an initial undirected bootstrap dataset size on performance, finding that a modest amount of unstructured demonstration data provided up-front drastically speeds up the convergence of a general purpose goal-reaching policy. See supplementary video for examples of the policy on a physical robot.

We propose a model-free algorithm for learning efficient policies capable of returning table tennis balls by controlling robot joints at a rate of 100Hz. We demonstrate that evolutionary search (ES) methods acting on CNN-based policy architectures for non-visual inputs and convolving across time learn compact controllers leading to smooth motions. Furthermore, we show that with appropriately tuned curriculum learning on the task and rewards, policies are capable of developing multi-modal styles, specifically forehand and backhand stroke, whilst achieving 80% return rate on a wide range of ball throws. We observe that multi-modality does not require any architectural priors, such as multi-head architectures or hierarchical policies.

Laura Harding Graesser, Kyunghyun Cho, Douwe Kiela. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.