With the integration of large language models (LLMs), embodied agents have strong capabilities to execute complicated instructions in natural language, paving a way for the potential deployment of embodied robots. However, a foreseeable issue is that those embodied agents can also flawlessly execute some hazardous tasks, potentially causing damages in real world. To study this issue, we present SafeAgentBench -- a new benchmark for safety-aware task planning of embodied LLM agents. SafeAgentBench includes: (1) a new dataset with 750 tasks, covering 10 potential hazards and 3 task types; (2) SafeAgentEnv, a universal embodied environment with a low-level controller, supporting multi-agent execution with 17 high-level actions for 8 state-of-the-art baselines; and (3) reliable evaluation methods from both execution and semantic perspectives. Experimental results show that the best-performing baseline gets 69% success rate for safe tasks, but only 5% rejection rate for hazardous tasks, indicating significant safety risks. More details and codes are available at https://github.com/shengyin1224/SafeAgentBench.

This paper presents SibylSat, a novel SAT-based method designed to efficiently solve totally-ordered HTN problems (TOHTN). In contrast to prevailing SAT-based HTN planners that employ a breadth-first search strategy, SibylSat adopts a greedy search approach, enabling it to identify promising decompositions for expansion. The selection process is facilitated by a heuristic derived from solving a relaxed problem, which is also expressed as a SAT problem. Our experimental evaluations demonstrate that SibylSat outperforms existing SAT-based TOHTN approaches in terms of both runtime and plan quality on most of the IPC benchmarks, while also solving a larger number of problems.

We advance the field of Parameter-Efficient Fine-Tuning (PEFT) with our novel multi-adapter method, OrchMoE, which capitalizes on modular skill architecture for enhanced forward transfer in neural networks. Unlike prior models that depend on explicit task identification inputs, OrchMoE automatically discerns task categories, streamlining the learning process. This is achieved through an integrated mechanism comprising an Automatic Task Classification module and a Task-Skill Allocation module, which collectively deduce task-specific classifications and tailor skill allocation matrices. Our extensive evaluations on the 'Super Natural Instructions' dataset, featuring 1,600 diverse instructional tasks, indicate that OrchMoE substantially outperforms comparable multi-adapter baselines in terms of both performance and sample utilization efficiency, all while operating within the same parameter constraints. These findings suggest that OrchMoE offers a significant leap forward in multi-task learning efficiency.

Exploring sparse reward multi-agent reinforcement learning (MARL) environments with traps in a collaborative manner is a complex task. Agents typically fail to reach the goal state and fall into traps, which affects the overall performance of the system. To overcome this issue, we present SOMARL, a framework that uses prior knowledge to reduce the exploration space and assist learning. In SOMARL, agents are treated as part of the MARL environment, and symbolic knowledge is embedded using a tree structure to build a knowledge hierarchy. The framework has a two-layer hierarchical structure, comprising a hybrid module with a Hierarchical Task Network (HTN) planning and meta-controller at the higher level, and a MARL-based interactive module at the lower level. The HTN module and meta-controller use Hierarchical Domain Definition Language (HDDL) and the option framework to formalize symbolic knowledge and obtain domain knowledge and a symbolic option set, respectively. Moreover, the HTN module leverages domain knowledge to guide low-level agent exploration by assisting the meta-controller in selecting symbolic options. The meta-controller further computes intrinsic rewards of symbolic options to limit exploration behavior and adjust HTN planning solutions as needed. We evaluate SOMARL on two benchmarks, FindTreasure and MoveBox, and report superior performance over state-of-the-art MARL and subgoal-based baselines for MARL environments significantly.

The ability to acquire abstract knowledge is a hallmark of human intelligence and is believed by many to be one of the core differences between humans and neural network models. Agents can be endowed with an inductive bias towards abstraction through meta-learning, where they are trained on a distribution of tasks that share some abstract structure that can be learned and applied. However, because neural networks are hard to interpret, it can be difficult to tell whether agents have learned the underlying abstraction, or alternatively statistical patterns that are characteristic of that abstraction. In this work, we compare the performance of humans and agents in a meta-reinforcement learning paradigm in which tasks are generated from abstract rules. We define a novel methodology for building "task metamers" that closely match the statistics of the abstract tasks but use a different underlying generative process, and evaluate performance on both abstract and metamer tasks. We find that humans perform better at abstract tasks than metamer tasks whereas common neural network architectures typically perform worse on the abstract tasks than the matched metamers. This work provides a foundation for characterizing differences between humans and machine learning that can be used in future work towards developing machines with more human-like behavior.

Hierarchical planning has attracted renewed interest in the last couple of years. As a consequence, the time was right to establish a workshop devoted entirely to hierarchical planning – an insight shared by many supporters. In this paper we report on the first ICAPS workshop on Hierarchical Planning held in Delft, The Netherlands, in 2018 as well as on the second workshop held in Berkeley, CA, USA, in 2019. Hierarchical planning approaches incorporate hierarchies in the domain model. In the most common form, the hierarchy is defined among tasks, leading to the distinction between primitive and abstract tasks.

The majority of search-based HTN planning systems can be divided into those searching a space of partial plans (a plan space) and those performing progression search, i.e., that build the solution in a forward manner. So far, all HTN planners that guide the search by using heuristic functions are based on plan space search. Those systems represent the set of search nodes more effectively by maintaining a partial ordering between tasks, but they have only limited information about the current state during search. In this article, we propose the use of progression search as basis for heuristic HTN planning systems. Such systems can calculate their heuristics incorporating the current state, because it is tracked during search. Our contribution is the following: We introduce two novel progression algorithms that avoid unnecessary branching when the problem at hand is partially ordered and show that both are sound and complete. We show that defining systematicity is problematic for search in HTN planning, propose a definition, and show that it is fulfilled by one of our algorithms. Then, we introduce a method to apply arbitrary classical planning heuristics to guide the search in HTN planning. It relaxes the HTN planning model to a classical model that is only used for calculating heuristics. It is updated during search and used to create heuristic values that are used to guide the HTN search. We show that it can be used to create HTN heuristics with interesting theoretical properties like safety, goal-awareness, and admissibility. Our empirical evaluation shows that the resulting system outperforms the state of the art in search-based HTN planning.

The research in hierarchical planning has made considerable progress in the last few years. Many recent systems do not rely on hand-tailored advice anymore to find solutions, but are supposed to be domain-independent systems that come with sophisticated solving techniques. In principle, this development would make the comparison between systems easier (because the domains are not tailored to a single system anymore) and -- much more important -- also the integration into other systems, because the modeling process is less tedious (due to the lack of advice) and there is no (or less) commitment to a certain planning system the model is created for. However, these advantages are destroyed by the lack of a common input language and feature set supported by the different systems. In this paper, we propose an extension to PDDL, the description language used in non-hierarchical planning, to the needs of hierarchical planning systems. We restrict our language to a basic feature set shared by many recent systems, give an extension of PDDL's EBNF syntax definition, and discuss our extensions with respect to several planner-specific input languages from related work.

Over the last year, the amount of research in hierarchical planning has increased, leading to significant improvements in the performance of planners. However, the research is diverging and planners are somewhat hard to compare against each other. This is mostly caused by the fact that there is no standard set of benchmark domains, nor even a common description language for hierarchical planning problems. As a consequence, the available planners support a widely varying set of features and (almost) none of them can solve (or even parse) any problem developed for another planner. With this paper, we propose to create a new track for the IPC in which hierarchical planners will compete. This competition will result in a standardised description language, broader support for core features of that language among planners, a set of benchmark problems, a means to fairly and objectively compare HTN planners, and for new challenges for planners. Introduction When the International Planning Competition (IPC) started out in 1998 it aimed to include both classical and hierarchical planners as competitors (McDermott 2000).

In this paper, we propose a novel SAT-based planning approach for hierarchical planning by introducing the SAT-based planner totSAT for the class of totally-ordered HTN planning problems. We use the same general approach as SAT planning for classical planning does: bound the problem, translate the problem into a formula, and if the formula is not satisfiable, increase the bound. In HTN planning, a suitable bound is the maximum depth of decomposition. We show how totally-ordered HTN planning problems can be translated into a SAT formula, given this bound. Furthermore, we have conducted an extensive empirical evaluation to compare our new planner against state-of-the-art HTN planners. It shows that our technique outperforms any of these systems.