Reinforcement Learning from Human Feedback (RLHF) is currently the most widely used method to align large language models (LLMs) with human preferences. Existing RLHF methods can be roughly categorized as either reward-based or reward-free. Novel applications such as ChatGPT and Claude leverage reward-based methods that first learn a reward model and apply actor-critic algorithms, such as Proximal Policy Optimization (PPO). However, in academic benchmarks, state-of-the-art results are often achieved via reward-free methods, such as Direct Preference Optimization (DPO). Is DPO truly superior to PPO? Why does PPO perform poorly on these benchmarks? In this paper, we first conduct both theoretical and empirical studies on the algorithmic properties of DPO and show that DPO may have fundamental limitations. Moreover, we also comprehensively examine PPO and reveal the key factors for the best performances of PPO in fine-tuning LLMs. Finally, we benchmark DPO and PPO across various a collection of RLHF testbeds, ranging from dialogue to code generation. Experiment results demonstrate that PPO is able to surpass other alignment methods in all cases and achieve state-of-the-art results in challenging code competitions.
We present a large language model (LLM) based system to empower quadrupedal robots with problem-solving abilities for long-horizon tasks beyond short-term motions. Long-horizon tasks for quadrupeds are challenging since they require both a high-level understanding of the semantics of the problem for task planning and a broad range of locomotion and manipulation skills to interact with the environment. Our system builds a high-level reasoning layer with large language models, which generates hybrid discrete-continuous plans as robot code from task descriptions. It comprises multiple LLM agents: a semantic planner for sketching a plan, a parameter calculator for predicting arguments in the plan, and a code generator to convert the plan into executable robot code. At the low level, we adopt reinforcement learning to train a set of motion planning and control skills to unleash the flexibility of quadrupeds for rich environment interactions. Our system is tested on long-horizon tasks that are infeasible to complete with one single skill. Simulation and real-world experiments show that it successfully figures out multi-step strategies and demonstrates non-trivial behaviors, including building tools or notifying a human for help.
In this paper, we study the cooperative Multi-Agent Reinforcement Learning (MARL) problems using Reward Machines (RMs) to specify the reward functions such that the prior knowledge of high-level events in a task can be leveraged to facilitate the learning efficiency. Unlike the existing work that RMs have been incorporated into MARL for task decomposition and policy learning in relatively simple domains or with an assumption of independencies among the agents, we present Multi-Agent Reinforcement Learning with a Hierarchy of RMs (MAHRM) that is capable of dealing with more complex scenarios when the events among agents can occur concurrently and the agents are highly interdependent. MAHRM exploits the relationship of high-level events to decompose a task into a hierarchy of simpler subtasks that are assigned to a small group of agents, so as to reduce the overall computational complexity. Experimental results in three cooperative MARL domains show that MAHRM outperforms other MARL methods using the same prior knowledge of high-level events.
Primal-dual safe RL methods commonly perform iterations between the primal update of the policy and the dual update of the Lagrange Multiplier. Such a training paradigm is highly susceptible to the error in cumulative cost estimation since this estimation serves as the key bond connecting the primal and dual update processes. We show that this problem causes significant underestimation of cost when using off-policy methods, leading to the failure to satisfy the safety constraint. To address this issue, we propose \textit{conservative policy optimization}, which learns a policy in a constraint-satisfying area by considering the uncertainty in cost estimation. This improves constraint satisfaction but also potentially hinders reward maximization. We then introduce \textit{local policy convexification} to help eliminate such suboptimality by gradually reducing the estimation uncertainty. We provide theoretical interpretations of the joint coupling effect of these two ingredients and further verify them by extensive experiments. Results on benchmark tasks show that our method not only achieves an asymptotic performance comparable to state-of-the-art on-policy methods while using much fewer samples, but also significantly reduces constraint violation during training. Our code is available at https://github.com/ZifanWu/CAL.
AI agents powered by Large Language Models (LLMs) have made significant advances, enabling them to assist humans in diverse complex tasks and leading to a revolution in human-AI coordination. LLM-powered agents typically require invoking LLM APIs and employing artificially designed complex prompts, which results in high inference latency. While this paradigm works well in scenarios with minimal interactive demands, such as code generation, it is unsuitable for highly interactive and real-time applications, such as gaming. Traditional gaming AI often employs small models or reactive policies, enabling fast inference but offering limited task completion and interaction abilities. In this work, we consider Overcooked as our testbed where players could communicate with natural language and cooperate to serve orders. We propose a Hierarchical Language Agent (HLA) for human-AI coordination that provides both strong reasoning abilities while keeping real-time execution. In particular, HLA adopts a hierarchical framework and comprises three modules: a proficient LLM, referred to as Slow Mind, for intention reasoning and language interaction, a lightweight LLM, referred to as Fast Mind, for generating macro actions, and a reactive policy, referred to as Executor, for transforming macro actions into atomic actions. Human studies show that HLA outperforms other baseline agents, including slow-mind-only agents and fast-mind-only agents, with stronger cooperation abilities, faster responses, and more consistent language communications.
In this paper, we aim to utilize only offline trajectory data to train a policy for multi-objective RL. We extend the offline policy-regularized method, a widely-adopted approach for single-objective offline RL problems, into the multi-objective setting in order to achieve the above goal. However, such methods face a new challenge in offline MORL settings, namely the preference-inconsistent demonstration problem. We propose two solutions to this problem: 1) filtering out preference-inconsistent demonstrations via approximating behavior preferences, and 2) adopting regularization techniques with high policy expressiveness. Moreover, we integrate the preference-conditioned scalarized update method into policy-regularized offline RL, in order to simultaneously learn a set of policies using a single policy network, thus reducing the computational cost induced by the training of a large number of individual policies for various preferences. Finally, we introduce Regularization Weight Adaptation to dynamically determine appropriate regularization weights for arbitrary target preferences during deployment. Empirical results on various multi-objective datasets demonstrate the capability of our approach in solving offline MORL problems.
This paper addresses the problem of multi-agent pursuit, where slow pursuers cooperate to capture fast evaders in a confined environment with obstacles. Existing heuristic algorithms often lack expressive coordination strategies and are highly sensitive to task conditions, requiring extensive hyperparameter tuning. In contrast, reinforcement learning (RL) has been applied to this problem and is capable of obtaining cooperative pursuit strategies. However, RL-based methods face challenges in training for complex scenarios due to the vast amount of training data and limited adaptability to varying task conditions, such as different scene sizes, varying numbers and speeds of obstacles, and flexible speed ratios of the evader to the pursuer. In this work, we combine RL and curriculum learning to introduce a flexible solver for multiagent pursuit problems, named TaskFlex Solver (TFS), which is capable of solving multi-agent pursuit problems with diverse and dynamically changing task conditions in both 2-dimensional and 3-dimensional scenarios. TFS utilizes a curriculum learning method that constructs task distributions based on training progress, enhancing training efficiency and final performance. Our algorithm consists of two main components: the Task Evaluator, which evaluates task success rates and selects tasks of moderate difficulty to maintain a curriculum archive, and the Task Sampler, which constructs training distributions by sampling tasks from the curriculum archive to maximize policy improvement. Experiments show that TFS produces much stronger performance than baselines and achieves close to 100% capture rates in both 2-dimensional and 3-dimensional multi-agent pursuit problems with diverse and dynamically changing scenes. The project website is at https://sites.google.com/view/tfs-2023.
We investigate the problem of decentralized multi-agent navigation tasks, where multiple agents need to reach initially unassigned targets in a limited time. Classical planning-based methods suffer from expensive computation overhead at each step and offer limited expressiveness for complex cooperation strategies. In contrast, reinforcement learning (RL) has recently become a popular paradigm for addressing this issue. However, RL struggles with low data efficiency and cooperation when directly exploring (nearly) optimal policies in the large search space, especially with an increased agent number (e.g., 10+ agents) or in complex environments (e.g., 3D simulators). In this paper, we propose Multi-Agent Scalable GNN-based P lanner (MASP), a goal-conditioned hierarchical planner for navigation tasks with a substantial number of agents. MASP adopts a hierarchical framework to divide a large search space into multiple smaller spaces, thereby reducing the space complexity and accelerating training convergence. We also leverage graph neural networks (GNN) to model the interaction between agents and goals, improving goal achievement. Besides, to enhance generalization capabilities in scenarios with unseen team sizes, we divide agents into multiple groups, each with a previously trained number of agents. The results demonstrate that MASP outperforms classical planning-based competitors and RL baselines, achieving a nearly 100% success rate with minimal training data in both multi-agent particle environments (MPE) with 50 agents and a quadrotor 3-dimensional environment (OmniDrones) with 20 agents. Furthermore, the learned policy showcases zero-shot generalization across unseen team sizes.
In this paper, a computer-vision-assisted simulation method is proposed to address the issue of training dataset acquisition for wireless hand gesture recognition. In the existing literature, in order to classify gestures via the wireless channel estimation, massive training samples should be measured in a consistent environment, consuming significant efforts. In the proposed CASTER simulator, however, the training dataset can be simulated via existing videos. Particularly, a gesture is represented by a sequence of snapshots, and the channel impulse response of each snapshot is calculated via tracing the rays scattered off a primitive-based hand model. Moreover, CASTER simulator relies on the existing videos to extract the motion data of gestures. Thus, the massive measurements of wireless channel can be eliminated. The experiments demonstrate a 90.8% average classification accuracy of simulation-to-reality inference.