The use of preference annotated data for training machine learning models has a long history going back to early algorithms for recommender systems and market research (Bonilla et al., 2010; Boutilier, 2002; Guo and Sanner, 2010). These days preference optimization algorithms are receiving renewed attention since they are a natural candidate for shaping the outputs of deep learning systems, such as large language models (Ouyang et al., 2022; Team et al., 2024) or control policies, via human feedback (Azar et al., 2023; Christiano et al., 2017; Rafailov et al., 2023). Arguably, preference optimization algorithms can also be a natural choice even when direct human feedback is not available but one instead aims to optimize a machine learning model based on feedback from a hand-coded or learned critic function (judging desirability of solutions). Here preference optimization methods are useful since they let us optimize the model to achieve desired outcomes based on relative rankings between outcomes alone (rather than requiring absolute labels or carefully crafted reward functions). Among preference optimization approaches, those based on directly using preference data - as opposed to casting preference optimization as reinforcement learning from (human) feedback - such as DPO (Rafailov et al., 2023), have emerged as particularly successful since they only require access to an offline dataset of paired preference data, and are fairly robust to application domain and hyperparameter settings. However, algorithms within this class make specific assumptions tailored to their application domain. They were designed to optimize LLMs from human feedback in the form of comparisons of generated sentences and thus, by design, require paired preference data (since they directly model a specific choice of preference distribution). We are interested in finding algorithms that are more flexible, and applicable in settings where the assumptions underlying DPO do not apply.

Recent advances in real-world applications of reinforcement learning (RL) have relied on the ability to accurately simulate systems at scale. However, domains such as fluid dynamical systems exhibit complex dynamic phenomena that are hard to simulate at high integration rates, limiting the direct application of modern deep RL algorithms to often expensive or safety critical hardware. In this work, we introduce "Box o Flows", a novel benchtop experimental control system for systematically evaluating RL algorithms in dynamic real-world scenarios. We describe the key components of the Box o Flows, and through a series of experiments demonstrate how state-of-the-art model-free RL algorithms can synthesize a variety of complex behaviors via simple reward specifications. Furthermore, we explore the role of offline RL in data-efficient hypothesis testing by reusing past experiences. We believe that the insights gained from this preliminary study and the availability of systems like the Box o Flows support the way forward for developing systematic RL algorithms that can be generally applied to complex, dynamical systems. Supplementary material and videos of experiments are available at https://sites.google.com/view/box-o-flows/home.

We show that offline actor-critic reinforcement learning can scale to large models - such as transformers - and follows similar scaling laws as supervised learning. We find that offline actor-critic algorithms can outperform strong, supervised, behavioral cloning baselines for multi-task training on a large dataset containing both sub-optimal and expert behavior on 132 continuous control tasks. We introduce a Perceiver-based actor-critic model and elucidate the key model features needed to make offline RL work with self- and cross-attention modules. Overall, we find that: i) simple offline actor critic algorithms are a natural choice for gradually moving away from the currently predominant paradigm of behavioral cloning, and ii) via offline RL it is possible to learn multi-task policies that master many domains simultaneously, including real robotics tasks, from sub-optimal demonstrations or self-generated data.

The ability to leverage heterogeneous robotic experience from different robots and tasks to quickly master novel skills and embodiments has the potential to transform robot learning. Inspired by recent advances in foundation models for vision and language, we propose a multi-embodiment, multi-task generalist agent for robotic manipulation. This agent, named RoboCat, is a visual goal-conditioned decision transformer capable of consuming action-labelled visual experience. This data spans a large repertoire of motor control skills from simulated and real robotic arms with varying sets of observations and actions. With RoboCat, we demonstrate the ability to generalise to new tasks and robots, both zero-shot as well as through adaptation using only 100-1000 examples for the target task. We also show how a trained model itself can be used to generate data for subsequent training iterations, thus providing a basic building block for an autonomous improvement loop. We investigate the agent's capabilities, with large-scale evaluations both in simulation and on three different real robot embodiments. We find that as we grow and diversify its training data, RoboCat not only shows signs of cross-task transfer, but also becomes more efficient at adapting to new tasks.

Reinforcement learning solely from an agent's self-generated data is often believed to be infeasible for learning on real robots, due to the amount of data needed. However, if done right, agents learning from real data can be surprisingly efficient through re-using previously collected sub-optimal data. In this paper we demonstrate how the increased understanding of off-policy learning methods and their embedding in an iterative online/offline scheme (``collect and infer'') can drastically improve data-efficiency by using all the collected experience, which empowers learning from real robot experience only. Moreover, the resulting policy improves significantly over the state of the art on a recently proposed real robot manipulation benchmark. Our approach learns end-to-end, directly from pixels, and does not rely on additional human domain knowledge such as a simulator or demonstrations.

We enable reinforcement learning agents to learn successful behavior policies by utilizing relevant pre-existing teacher policies. The teacher policies are introduced as objectives, in addition to the task objective, in a multi-objective policy optimization setting. Using the Multi-Objective Maximum a Posteriori Policy Optimization algorithm (Abdolmaleki et al. 2020), we show that teacher policies can help speed up learning, particularly in the absence of shaping rewards. In two domains with continuous observation and action spaces, our agents successfully compose teacher policies in sequence and in parallel, and are also able to further extend the policies of the teachers in order to solve the task. Depending on the specified combination of task and teacher(s), teacher(s) may naturally act to limit the final performance of an agent. The extent to which agents are required to adhere to teacher policies are determined by hyperparameters which determine both the effect of teachers on learning speed and the eventual performance of the agent on the task. In the humanoid domain (Tassa et al. 2018), we also equip agents with the ability to control the selection of teachers. With this ability, agents are able to meaningfully compose from the teacher policies to achieve a superior task reward on the walk task than in cases without access to the teacher policies. We show the resemblance of composed task policies with the corresponding teacher policies through videos.

Many advances that have improved the robustness and efficiency of deep reinforcement learning (RL) algorithms can, in one way or another, be understood as introducing additional objectives or constraints in the policy optimization step. This includes ideas as far ranging as exploration bonuses, entropy regularization, and regularization toward teachers or data priors. Often, the task reward and auxiliary objectives are in conflict, and in this paper we argue that this makes it natural to treat these cases as instances of multi-objective (MO) optimization problems. We demonstrate how this perspective allows us to develop novel and more effective RL algorithms. In particular, we focus on offline RL and finetuning as case studies, and show that existing approaches can be understood as MO algorithms relying on linear scalarization. We hypothesize that replacing linear scalarization with a better algorithm can improve performance. We introduce Distillation of a Mixture of Experts (DiME), a new MORL algorithm that outperforms linear scalarization and can be applied to these non-standard MO problems. We demonstrate that for offline RL, DiME leads to a simple new algorithm that outperforms state-of-the-art. For finetuning, we derive new algorithms that learn to outperform the teacher policy. Deep reinforcement learning (RL) algorithms have solved a number of challenging problems, including in games (Mnih et al., 2015; Silver et al., 2016), simulated continuous control (Heess et al., 2017; Peng et al., 2018), and robotics (OpenAI et al., 2018). The standard RL setting appeals through its simplicity: an agent acts in the environment and can discover complex solutions simply by maximizing cumulative discounted reward. In practice, however, the situation is often more complicated. For instance, without a carefully crafted reward function or sophisticated exploration strategy, learning may require hundreds of millions of environment interactions, or may not be possible at all. A number of strategies have been developed to mitigate the shortcomings of the pure RL paradigm. These include strategies that regularize the final solution, for instance by maximizing auxiliary rewards (Jaderberg et al., 2017) or the entropy of the policy (Mnih et al., 2016; Haarnoja et al., 2018).

From daily interactions with the world, humans gradually develop an internal understanding of which series of events would be triggered when a certain sequence of actions is taken (Hogendoorn and Burkitt, 2018; Maus et al., 2013; Nortmann et al., 2015). This mental model of the world can serve as a compact proxy of our previous experiences and help us plan out routes to desired goals before taking action (Ha and Schmidhuber, 2018). Studies have further implied that these mental predictive models might not be restricted to the level of primitive actions (Botvinick, 2008; Consul et al., 2022), but rather consider predictions over larger timescales that abstract away detailed behavior consequences, which can enable efficient long-horizon planning to guide our daily decision making. When developing intelligent artificial agents it is therefore natural to imagine a similar process being useful for learning and transferring abstract models of the world across streams of experiences and tasks. We expect such a temporally abstract model of actions and dynamics to be significantly more useful than a simple one-step prediction model (together with primitive policies) when transferring them to a target task. This is because they should allow us to rapidly plan over long trajectories (to find some states with high rewards) while alleviating the common problem of error accumulation that occurs when chaining one-step prediction models which limits the effective planning horizon in most existing methods, e.g.

The ability to effectively reuse prior knowledge is a key requirement when building general and flexible Reinforcement Learning (RL) agents. Skill reuse is one of the most common approaches, but current methods have considerable limitations.For example, fine-tuning an existing policy frequently fails, as the policy can degrade rapidly early in training. In a similar vein, distillation of expert behavior can lead to poor results when given sub-optimal experts. We compare several common approaches for skill transfer on multiple domains including changes in task and system dynamics. We identify how existing methods can fail and introduce an alternative approach to mitigate these problems. Our approach learns to sequence existing temporally-extended skills for exploration but learns the final policy directly from the raw experience. This conceptual split enables rapid adaptation and thus efficient data collection but without constraining the final solution.It significantly outperforms many classical methods across a suite of evaluation tasks and we use a broad set of ablations to highlight the importance of differentc omponents of our method.

There is a widespread intuition that model-based control methods should be able to surpass the data efficiency of model-free approaches. In this paper we attempt to evaluate this intuition on various challenging locomotion tasks. We take a hybrid approach, combining model predictive control (MPC) with a learned model and model-free policy learning; the learned policy serves as a proposal for MPC. We find that well-tuned model-free agents are strong baselines even for high DoF control problems but MPC with learned proposals and models (trained on the fly or transferred from related tasks) can significantly improve performance and data efficiency in hard multi-task/multi-goal settings. Finally, we show that it is possible to distil a model-based planner into a policy that amortizes the planning computation without any loss of performance. Videos of agents performing different tasks can be seen at https://sites.google.com/view/mbrl-amortization/home.