Long-horizon embodied planning is challenging because the world does not only change through an agent's actions: exogenous processes (e.g., water heating, dominoes cascading) unfold concurrently with the agent's actions. We propose a framework for abstract world models that jointly learns (i) symbolic state representations and (ii) causal processes for both endogenous actions and exogenous mechanisms. Each causal process models the time course of a stochastic cause-effect relation. We learn these world models from limited data via variational Bayesian inference combined with LLM proposals. Across five simulated tabletop robotics environments, the learned models enable fast planning that generalizes to held-out tasks with more objects and more complex goals, outperforming a range of baselines.

Our aim is to learn to solve long-horizon decision-making problems in highly-variable, combinatorially-complex robotics domains given raw sensor input in the form of images. Previous work has shown that one way to achieve this aim is to learn a structured abstract transition model in the form of symbolic predicates and operators, and then plan within this model to solve novel tasks at test time. However, these learned models do not ground directly into pixels from just a handful of demonstrations. In this work, we propose to invent predicates that operate directly over input images by leveraging the capabilities of pretrained vision-language models (VLMs). Our key idea is that, given a set of demonstrations, a VLM can be used to propose a set of predicates that are potentially relevant for decision-making and then to determine the truth values of these predicates in both the given demonstrations and new image inputs. We build upon an existing framework for predicate invention, which generates feature-based predicates operating on object-centric states, to also generate visual predicates that operate on images. Experimentally, we show that our approach -- pix2pred -- is able to invent semantically meaningful predicates that enable generalization to novel, complex, and long-horizon tasks across two simulated robotic environments.

Broadly intelligent agents should form task-specific abstractions that selectively expose the essential elements of a task, while abstracting away the complexity of the raw sensorimotor space. In this work, we present Neuro-Symbolic Predicates, a first-order abstraction language that combines the strengths of symbolic and neural knowledge representations. We outline an online algorithm for inventing such predicates and learning abstract world models. We compare our approach to hierarchical reinforcement learning, vision-language model planning, and symbolic predicate invention approaches, on both in- and out-of-distribution tasks across five simulated robotic domains. Results show that our approach offers better sample complexity, stronger out-of-distribution generalization, and improved interpretability.

Previous STRIPS domain model acquisition approaches that learn from state traces start with the names and parameters of the actions to be learned. Therefore their only task is to deduce the preconditions and effects of the given actions. In this work, we explore learning in situations when the parameters of learned actions are not provided. We define two levels of trace quality based on which information is provided and present an algorithm for each. In one level (L1), the states in the traces are labeled with action names, so we can deduce the number and names of the actions, but we still need to work out the number and types of parameters. In the other level (L2), the states are additionally labeled with objects that constitute the parameters of the corresponding grounded actions. Here we still need to deduce the types of the parameters in the learned actions. We experimentally evaluate the proposed algorithms and compare them with the state-of-the-art learning tool FAMA on a large collection of IPC benchmarks. The evaluation shows that our new algorithms are faster, can handle larger inputs and provide better results in terms of learning action models more similar to reference models.

Abstract: An effective approach to solving long-horizon tasks in robotics domains with continuous state and action spaces is bilevel planning, wherein a highlevel search over an abstraction of an environment is used to guide low-level decision-making. Recent work has shown how to enable such bilevel planning by learning abstract models in the form of symbolic operators and neural samplers. In this work, we show that existing symbolic operator learning approaches fall short in many robotics domains where a robot's actions tend to cause a large number of irrelevant changes in the abstract state. This is primarily because they attempt to learn operators that exactly predict all observed changes in the abstract state. To overcome this issue, we propose to learn operators that'choose what to predict' by only modelling changes necessary for abstract planning to achieve specified goals. Experimentally, we show that our approach learns operators that lead to efficient planning across 10 different hybrid robotics domains, including 4 from the challenging BEHAVIOR-100 benchmark, while generalizing to novel initial states, goals, and objects.

Temporal planning is an extension of classical planning involving concurrent execution of actions and alignment with temporal constraints. Durative actions along with invariants allow for modeling domains in which multiple agents operate in parallel on shared resources. Hence, it is often important to avoid resource conflicts, where temporal constraints establish the consistency of concurrent actions and events. Unfortunately, the performance of temporal planning engines tends to sharply deteriorate when the number of agents and objects in a domain gets large. A possible remedy is to use macro-actions that are well-studied in the context of classical planning. In temporal planning settings, however, introducing macro-actions is significantly more challenging when the concurrent execution of actions and shared use of resources, provided the compliance to temporal constraints, should not be suppressed entirely. Our work contributes a general concept of sequential temporal macro-actions that guarantees the applicability of obtained plans, i.e., the sequence of original actions encapsulated by a macro-action is always executable. We apply our approach to several temporal planners and domains, stemming from the International Planning Competition and RoboCup Logistics League. Our experiments yield improvements in terms of obtained satisficing plans as well as plan quality for the majority of tested planners and domains.

We present a technique for automatically extracting temporal mutual exclusion invariants from PDDL2.2 planning instances. Our technique builds on other approaches to invariant synthesis presented in the literature, but departs from their limited focus on instantaneous discrete actions by addressing temporal and numeric domains. To deal with time, we formulate invariance conditions that account for both the entire structure of the operators (including the conditions, rather than just the effects) and the possible interactions between operators.

Between 1998 and 2004, the planning community has seen vast progress in terms of the sizes of benchmark examples that domain-independent planners can tackle successfully. The key technique behind this progress is the use of heuristic functions based on relaxing the planning task at hand, where the relaxation is to assume that all delete lists are empty. The unprecedented success of such methods, in many commonly used benchmark examples, calls for an understanding of what classes of domains these methods are well suited for. In the investigation at hand, we derive a formal background to such an understanding. We perform a case study covering a range of 30 commonly used STRIPS and ADL benchmark domains, including all examples used in the first four international planning competitions. We *prove* connections between domain structure and local search topology -- heuristic cost surface properties -- under an idealized version of the heuristic functions used in modern planners. The idealized heuristic function is called h^+, and differs from the practically used functions in that it returns the length of an *optimal* relaxed plan, which is NP-hard to compute. We identify several key characteristics of the topology under h^+, concerning the existence/non-existence of unrecognized dead ends, as well as the existence/non-existence of constant upper bounds on the difficulty of escaping local minima and benches. These distinctions divide the (set of all) planning domains into a taxonomy of classes of varying h^+ topology. As it turns out, many of the 30 investigated domains lie in classes with a relatively easy topology. Most particularly, 12 of the domains lie in classes where FFs search algorithm, provided with h^+, is a polynomial solving mechanism. We also present results relating h^+ to its approximation as implemented in FF. The behavior regarding dead ends is provably the same. We summarize the results of an empirical investigation showing that, in many domains, the topological qualities of h^+ are largely inherited by the approximation. The overall investigation gives a rare example of a successful analysis of the connections between typical-case problem structure, and search performance. The theoretical investigation also gives hints on how the topological phenomena might be automatically recognizable by domain analysis techniques. We outline some preliminary steps we made into that direction.

Engineering complete planning domain descriptions is often very costly because of human error or lack of domain knowl- edge. Learning complete domain descriptions is also very challenging because many features are irrelevant to achieving the goals and data may be scarce. We present a planner and agent that respectively plan and act in incomplete domains by i) synthesizing plans to avoid execution failure due to ignorance of the domain model, and ii) passively learning about the domain model during execution to improve later re-planning attempts. Our planner DeFault is the first to reason about a domain’s incompleteness to avoid potential plan failure. DeFault computes failure explanations for each action and state in the plan and counts the number of interpretations of the incomplete domain where failure will occur. We show that DeFault performs best by counting prime implicants (failure diagnoses) rather than propositional models. Our agent Goalie learns about the preconditions and effects of incompletely-specified actions while monitoring its state and, in conjunction with DeFault plan failure explanations, can diagnose past and future action failures. We show that by reasoning about incompleteness (as opposed to ignoring it) Goalie fails and re-plans less and executes fewer actions.

Forward State Planning with Reachability Heuristics is arguably the most successful approach to Automated Planning up to date. In addition to an estimation of the distance to the goal, relaxed plans obtained with such heuristics provide the search with useful information such as helpful actions and look-ahead states. However, this information is extracted only from the beginning of the relaxed plan. In this paper, we propose using information extracted from the last actions in the relaxed plan to generate intermediate goals backwards. This allows us to use information from previous computations of the heuristic and reduce the depth of the search tree.