Recent work on large language models has demonstrated the use of model-free reinforcement learning (RL) to train reasoning-like capabilities. The emergence of "thinking" through model-free RL is interesting as thinking actions neither produce reward nor change the external world state to one where the agent is more likely to get reward. This paper seeks to build a domain-independent understanding of when model-free RL will lead to such "thinking" as a strategy for reward maximization. To build this understanding, we first introduce a theoretical model which we call a thought Markov decision process (MDP). Thought MDPs minimally extend the classical MDP model to include an abstract notion of thought state and thought action. Using the thought MDP model, we prove the importance of policy initialization in determining whether or not thinking emerges and show formally that thought actions are equivalent to the agent choosing to perform a step of policy improvement before continuing to act. We then show that open-source LLMs satisfy the conditions that our theory predicts are necessary for model-free RL to produce thinking-like behavior. Finally, we hypothesize sufficient conditions that would enable thinking to be learned outside of language generation and introduce a toy domain where a combination of multi-task pre-training and designated thought actions enable more data-efficient RL compared to non-thinking agents.

Recent work on large language models has demonstrated the use of model-free reinforcement learning (RL) to train reasoning-like capabilities. The emergence of thinking through model-free RL is interesting as thinking actions neither produce reward nor change the external world state to one where the agent is more likely to get reward. This paper seeks to build a domain-independent understanding of when model-free RL will lead to such thinking as a strategy for reward maximization. To build this understanding, we first introduce a theoretical model which we call a thought Markov decision process (MDP). Thought MDPs minimally extend the classical MDP model to include an abstract notion of thought state and thought action. Using the thought MDP model, we prove the importance of policy initialization in determining whether or not thinking emerges and show formally that thought actions are equivalent to the agent choosing to perform a step of policy improvement before continuing to act. We then show that open-source LLMs satisfy the conditions that our theory predicts are necessary for model-free RL to produce thinking-like behavior. Finally, we hypothesize sufficient conditions that would enable thinking to be learned outside of language generation and introduce a toy domain where a combination of multi-task pre-training and designated thought actions enable more data-efficient RL compared to non-thinking agents.

Visual Reinforcement Learning (RL) methods often require extensive amounts of data. As opposed to model-free RL, model-based RL (MBRL) offers a potential solution with efficient data utilization through planning. Additionally, RL lacks generalization capabilities for real-world tasks. Prior work has shown that incorporating pre-trained visual representations (PVRs) enhances sample efficiency and generalization. While PVRs have been extensively studied in the context of model-free RL, their potential in MBRL remains largely unexplored.

Reinforcement learning (RL) is typically viewed as the problem of estimating single-step policies (for model-free RL) or single-step models (for model-based RL), leveraging the Markov property to factorize the problem in time. However, we can also view RL as a sequence modeling problem: predict a sequence of actions that leads to a sequence of high rewards. Viewed in this way, it is tempting to consider whether powerful, high-capacity sequence prediction models that work well in other supervised learning domains, such as natural-language processing, can also provide simple and effective solutions to the RL problem. To this end, we explore how RL can be reframed as one big sequence modeling problem, using state-of-the-art Transformer architectures to model distributions over sequences of states, actions, and rewards. Addressing RL as a sequence modeling problem significantly simplifies a range of design decisions: we no longer require separate behavior policy constraints, as is common in prior work on offline model-free RL, and we no longer require ensembles or other epistemic uncertainty estimators, as is common in prior work on model-based RL. All of these roles are filled by the same Transformer sequence model. In our experiments, we demonstrate the flexibility of this approach across imitation learning, goal-conditioned RL, and offline RL.

Recent work on large language models has demonstrated the use of model-free reinforcement learning (RL) to train reasoning-like capabilities. The emergence of "thinking" through model-free RL is interesting as thinking actions neither produce reward nor change the external world state to one where the agent is more likely to get reward. This paper seeks to build a domain-independent understanding of when model-free RL will lead to such "thinking" as a strategy for reward maximization. To build this understanding, we first introduce a theoretical model which we call a thought Markov decision process (MDP). Thought MDPs minimally extend the classical MDP model to include an abstract notion of thought state and thought action. Using the thought MDP model, we prove the importance of policy initialization in determining whether or not thinking emerges and show formally that thought actions are equivalent to the agent choosing to perform a step of policy improvement before continuing to act. We then show that open-source LLMs satisfy the conditions that our theory predicts are necessary for model-free RL to produce thinking-like behavior. Finally, we hypothesize sufficient conditions that would enable thinking to be learned outside of language generation and introduce a toy domain where a combination of multi-task pre-training and designated thought actions enable more data-efficient RL compared to non-thinking agents.

Visual Reinforcement Learning (RL) methods often require extensive amounts of data. As opposed to model-free RL, model-based RL (MBRL) offers a potential solution with efficient data utilization through planning. Additionally, RL lacks generalization capabilities for real-world tasks. Prior work has shown that incorporating pre-trained visual representations (PVRs) enhances sample efficiency and generalization. While PVRs have been extensively studied in the context of model-free RL, their potential in MBRL remains largely unexplored.

We propose an online motion planner for legged robot locomotion with the primary objective of achieving energy efficiency. The conceptual idea is to leverage a placement set of footstep positions based on the robot's body position to determine when and how to execute steps. In particular, the proposed planner uses virtual placement sets beneath the hip joints of the legs and executes a step when the foot is outside of such placement set. Furthermore, we propose a parameter design framework that considers both energy-efficiency and robustness measures to optimize the gait by changing the shape of the placement set along with other parameters, such as step height and swing time, as a function of walking speed. We show that the planner produces trajectories that have a low Cost of Transport (CoT) and high robustness measure, and evaluate our approach against model-free Reinforcement Learning (RL) and motion imitation using biological dog motion priors as the reference. Overall, within low to medium velocity range, we show a 50.4% improvement in CoT and improved robustness over model-free RL, our best performing baseline. Finally, we show ability to handle slippery surfaces, gait transitions, and disturbances in simulation and hardware with the Unitree A1 robot.

In this paper, we present an approach for learning collision-free robot trajectories in the presence of moving obstacles. As a first step, we train a backup policy to generate evasive movements from arbitrary initial robot states using model-free reinforcement learning. When learning policies for other tasks, the backup policy can be used to estimate the potential risk of a collision and to offer an alternative action if the estimated risk is considered too high. No matter which action is selected, our action space ensures that the kinematic limits of the robot joints are not violated. We analyze and evaluate two different methods for estimating the risk of a collision. A physics simulation performed in the background is computationally expensive but provides the best results in deterministic environments. If a data-based risk estimator is used instead, the computational effort is significantly reduced, but an additional source of error is introduced. For evaluation, we successfully learn a reaching task and a basketball task while keeping the risk of collisions low. The results demonstrate the effectiveness of our approach for deterministic and stochastic environments, including a human-robot scenario and a ball environment, where no state can be considered permanently safe. By conducting experiments with a real robot, we show that our approach can generate safe trajectories in real time.

Reinforcement learning (RL) is typically viewed as the problem of estimating single-step policies (for model-free RL) or single-step models (for model-based RL), leveraging the Markov property to factorize the problem in time. However, we can also view RL as a sequence modeling problem: predict a sequence of actions that leads to a sequence of high rewards. Viewed in this way, it is tempting to consider whether powerful, high-capacity sequence prediction models that work well in other supervised learning domains, such as natural-language processing, can also provide simple and effective solutions to the RL problem. To this end, we explore how RL can be reframed as "one big sequence modeling" problem, using state-of-the-art Transformer architectures to model distributions over sequences of states, actions, and rewards. Addressing RL as a sequence modeling problem significantly simplifies a range of design decisions: we no longer require separate behavior policy constraints, as is common in prior work on offline model-free RL, and we no longer require ensembles or other epistemic uncertainty estimators, as is common in prior work on model-based RL.

System optimal traffic routing can mitigate congestion by assigning routes for a portion of vehicles so that the total travel time of all vehicles in the transportation system can be reduced. However, achieving real-time optimal routing poses challenges due to uncertain demands and unknown system dynamics, particularly in expansive transportation networks. While physics model-based methods are sensitive to uncertainties and model mismatches, model-free reinforcement learning struggles with learning inefficiencies and interpretability issues. Our paper presents TransRL, a novel algorithm that integrates reinforcement learning with physics models for enhanced performance, reliability, and interpretability. TransRL begins by establishing a deterministic policy grounded in physics models, from which it learns from and is guided by a differentiable and stochastic teacher policy. During training, TransRL aims to maximize cumulative rewards while minimizing the Kullback Leibler (KL) divergence between the current policy and the teacher policy. This approach enables TransRL to simultaneously leverage interactions with the environment and insights from physics models. We conduct experiments on three transportation networks with up to hundreds of links. The results demonstrate TransRL's superiority over traffic model-based methods for being adaptive and learning from the actual network data. By leveraging the information from physics models, TransRL consistently outperforms state-of-the-art reinforcement learning algorithms such as proximal policy optimization (PPO) and soft actor critic (SAC). Moreover, TransRL's actions exhibit higher reliability and interpretability compared to baseline reinforcement learning approaches like PPO and SAC.