Position: Foundation Agents as the Paradigm Shift for Decision Making

Decision making demands intricate interplay between perception, memory, and reasoning to discern optimal policies. Conventional approaches to decision making face challenges related to low sample efficiency and poor generalization. In contrast, foundation models in language and vision have showcased rapid adaptation to diverse new tasks. Therefore, we advocate for the construction of foundation agents as a transformative shift in the learning paradigm of agents. This proposal is underpinned by the formulation of foundation agents with their fundamental characteristics and challenges motivated by the success of large language models (LLMs). Moreover, we specify the roadmap of foundation agents from large interactive data collection or generation, to self-supervised pretraining and adaptation, and knowledge and value alignment with LLMs. Lastly, we pinpoint critical research questions derived from the formulation and delineate trends for foundation agents supported by real-world use cases, addressing both technical and theoretical aspects to propel the field towards a more comprehensive and impactful future.

SPO: Multi-Dimensional Preference Sequential Alignment With Implicit Reward Modeling

Human preference alignment is critical in building powerful and reliable large language models (LLMs). However, current methods either ignore the multi-dimensionality of human preferences (e.g. helpfulness and harmlessness) or struggle with the complexity of managing multiple reward models. To address these issues, we propose Sequential Preference Optimization (SPO), a method that sequentially fine-tunes LLMs to align with multiple dimensions of human preferences. SPO avoids explicit reward modeling, directly optimizing the models to align with nuanced human preferences. We theoretically derive closed-form optimal SPO policy and loss function. Gradient analysis is conducted to show how SPO manages to fine-tune the LLMs while maintaining alignment on previously optimized dimensions. Empirical results on LLMs of different size and multiple evaluation datasets demonstrate that SPO successfully aligns LLMs across multiple dimensions of human preferences and significantly outperforms the baselines.

Urban Scene Diffusion through Semantic Occupancy Map

Generating unbounded 3D scenes is crucial for large-scale scene understanding and simulation. Urban scenes, unlike natural landscapes, consist of various complex man-made objects and structures such as roads, traffic signs, vehicles, and buildings. To create a realistic and detailed urban scene, it is crucial to accurately represent the geometry and semantics of the underlying objects, going beyond their visual appearance. In this work, we propose UrbanDiffusion, a 3D diffusion model that is conditioned on a Bird's-Eye View (BEV) map and generates an urban scene with geometry and semantics in the form of semantic occupancy map. Our model introduces a novel paradigm that learns the data distribution of scene-level structures within a latent space and further enables the expansion of the synthesized scene into an arbitrary scale. After training on real-world driving datasets, our model can generate a wide range of diverse urban scenes given the BEV maps from the held-out set and also generalize to the synthesized maps from a driving simulator. We further demonstrate its application to scene image synthesis with a pretrained image generator as a prior.

S-NeRF++: Autonomous Driving Simulation via Neural Reconstruction and Generation

Autonomous driving simulation system plays a crucial role in enhancing self-driving data and simulating complex and rare traffic scenarios, ensuring navigation safety. However, traditional simulation systems, which often heavily rely on manual modeling and 2D image editing, struggled with scaling to extensive scenes and generating realistic simulation data. In this study, we present S-NeRF++, an innovative autonomous driving simulation system based on neural reconstruction. Trained on widely-used self-driving datasets such as nuScenes and Waymo, S-NeRF++ can generate a large number of realistic street scenes and foreground objects with high rendering quality as well as offering considerable flexibility in manipulation and simulation. Specifically, S-NeRF++ is an enhanced neural radiance field for synthesizing large-scale scenes and moving vehicles, with improved scene parameterization and camera pose learning. The system effectively utilizes noisy and sparse LiDAR data to refine training and address depth outliers, ensuring high quality reconstruction and novel-view rendering. It also provides a diverse foreground asset bank through reconstructing and generating different foreground vehicles to support comprehensive scenario creation. Moreover, we have developed an advanced foreground-background fusion pipeline that skillfully integrates illumination and shadow effects, further enhancing the realism of our simulations. With the high-quality simulated data provided by our S-NeRF++, we found the perception methods enjoy performance boost on several autonomous driving downstream tasks, which further demonstrate the effectiveness of our proposed simulator.

Safe Reinforcement Learning with Free-form Natural Language Constraints and Pre-Trained Language Models

Safe reinforcement learning (RL) agents accomplish given tasks while adhering to specific constraints. Employing constraints expressed via easily-understandable human language offers considerable potential for real-world applications due to its accessibility and non-reliance on domain expertise. Previous safe RL methods with natural language constraints typically adopt a recurrent neural network, which leads to limited capabilities when dealing with various forms of human language input. Furthermore, these methods often require a ground-truth cost function, necessitating domain expertise for the conversion of language constraints into a well-defined cost function that determines constraint violation. To address these issues, we proposes to use pre-trained language models (LM) to facilitate RL agents' comprehension of natural language constraints and allow them to infer costs for safe policy learning. Through the use of pre-trained LMs and the elimination of the need for a ground-truth cost, our method enhances safe policy learning under a diverse set of human-derived free-form natural language constraints. Experiments on grid-world navigation and robot control show that the proposed method can achieve strong performance while adhering to given constraints. The usage of pre-trained LMs allows our method to comprehend complicated constraints and learn safe policies without the need for ground-truth cost at any stage of training or evaluation. Extensive ablation studies are conducted to demonstrate the efficacy of each part of our method.

TAPE: Leveraging Agent Topology for Cooperative Multi-Agent Policy Gradient

Multi-Agent Policy Gradient (MAPG) has made significant progress in recent years. However, centralized critics in state-of-the-art MAPG methods still face the centralized-decentralized mismatch (CDM) issue, which means sub-optimal actions by some agents will affect other agent's policy learning. While using individual critics for policy updates can avoid this issue, they severely limit cooperation among agents. To address this issue, we propose an agent topology framework, which decides whether other agents should be considered in policy gradient and achieves compromise between facilitating cooperation and alleviating the CDM issue. The agent topology allows agents to use coalition utility as learning objective instead of global utility by centralized critics or local utility by individual critics. To constitute the agent topology, various models are studied. We propose Topology-based multi-Agent Policy gradiEnt (TAPE) for both stochastic and deterministic MAPG methods. We prove the policy improvement theorem for stochastic TAPE and give a theoretical explanation for the improved cooperation among agents. Experiment results on several benchmarks show the agent topology is able to facilitate agent cooperation and alleviate CDM issue respectively to improve performance of TAPE. Finally, multiple ablation studies and a heuristic graph search algorithm are devised to show the efficacy of the agent topology.

Self-supervised Pretraining for Decision Foundation Model: Formulation, Pipeline and Challenges

Decision-making is a dynamic process requiring perception, memory, and reasoning to make choices and find optimal policies. Traditional approaches to decision-making suffer from sample efficiency and generalization, while large-scale self-supervised pretraining has enabled fast adaptation with fine-tuning or few-shot learning in language and vision. We thus argue to integrate knowledge acquired from generic large-scale self-supervised pretraining into downstream decision-making problems. We propose Pretrain-Then-Adapt pipeline and survey recent work on data collection, pretraining objectives and adaptation strategies for decision-making pretraining and downstream inference. Finally, we identify critical challenges and future directions for developing decision foundation model with the help of generic and flexible self-supervised pretraining.

PDiT: Interleaving Perception and Decision-making Transformers for Deep Reinforcement Learning

Designing better deep networks and better reinforcement learning (RL) algorithms are both important for deep RL. This work studies the former. Specifically, the Perception and Decision-making Interleaving Transformer (PDiT) network is proposed, which cascades two Transformers in a very natural way: the perceiving one focuses on \emph{the environmental perception} by processing the observation at the patch level, whereas the deciding one pays attention to \emph{the decision-making} by conditioning on the history of the desired returns, the perceiver's outputs, and the actions. Such a network design is generally applicable to a lot of deep RL settings, e.g., both the online and offline RL algorithms under environments with either image observations, proprioception observations, or hybrid image-language observations. Extensive experiments show that PDiT can not only achieve superior performance than strong baselines in different settings but also extract explainable feature representations. Our code is available at \url{https://github.com/maohangyu/PDiT}.

BadRL: Sparse Targeted Backdoor Attack Against Reinforcement Learning

Backdoor attacks in reinforcement learning (RL) have previously employed intense attack strategies to ensure attack success. However, these methods suffer from high attack costs and increased detectability. In this work, we propose a novel approach, BadRL, which focuses on conducting highly sparse backdoor poisoning efforts during training and testing while maintaining successful attacks. Our algorithm, BadRL, strategically chooses state observations with high attack values to inject triggers during training and testing, thereby reducing the chances of detection. In contrast to the previous methods that utilize sample-agnostic trigger patterns, BadRL dynamically generates distinct trigger patterns based on targeted state observations, thereby enhancing its effectiveness. Theoretical analysis shows that the targeted backdoor attack is always viable and remains stealthy under specific assumptions. Empirical results on various classic RL tasks illustrate that BadRL can substantially degrade the performance of a victim agent with minimal poisoning efforts 0.003% of total training steps) during training and infrequent attacks during testing.

Benchmarking Continual Learning from Cognitive Perspectives

Continual learning addresses the problem of continuously acquiring and transferring knowledge without catastrophic forgetting of old concepts. While humans achieve continual learning via diverse neurocognitive mechanisms, there is a mismatch between cognitive properties and evaluation methods of continual learning models. First, the measurement of continual learning models mostly relies on evaluation metrics at a micro-level, which cannot characterize cognitive capacities of the model. Second, the measurement is method-specific, emphasizing model strengths in one aspect while obscuring potential weaknesses in other respects. To address these issues, we propose to integrate model cognitive capacities and evaluation metrics into a unified evaluation paradigm. We first characterize model capacities via desiderata derived from cognitive properties supporting human continual learning. The desiderata concern (1) adaptability in varying lengths of task sequence; (2) sensitivity to dynamic task variations; and (3) efficiency in memory usage and training time consumption. Then we design evaluation protocols for each desideratum to assess cognitive capacities of recent continual learning models. Experimental results show that no method we consider has satisfied all the desiderata and is still far away from realizing truly continual learning. Although some methods exhibit some degree of adaptability and efficiency, no method is able to identify task relationships when encountering dynamic task variations, or achieve a trade-off in learning similarities and differences between tasks. Inspired by these results, we discuss possible factors that influence model performance in these desiderata and provide guidance for the improvement of continual learning models.