Classical symbolic planners, such as those utilizing the Planning Domain Definition Language (PDDL), excel at solving long-horizon tasks because they provide rigorous logical guarantees. By breaking down a complex goal into a sequence of discrete actions, these systems can reason over extended timelines far more reliably than pure end-to-end neural networks. However, as noted by Jonathan Tremblay et al., this paradigm suffers from a severe limitation: it requires human experts to manually define the symbolic rules, logical predicates, and action operators for every possible interaction.
This reliance on hand-crafted abstractions creates what is widely known as the knowledge engineering bottleneck. For a robot to plan how to clean a kitchen, a human engineer must explicitly define predicates like On(Object, Surface) or Holding(Robot, Object), alongside the exact preconditions and effects for every action. As Naman Shah et al. point out, this manual translation prevents robots from learning new tasks independently and severely limits their deployment in unstructured or novel environments. To achieve true autonomy, systems must transition from human-defined symbolic logic to fully autonomous predicate discovery, an open challenge highlighted by Alexandre Bernardino et al..
Why Manual Modeling Fails to Scale
To understand why manual domain modeling fails, consider the interaction between high-level task planning and low-level geometric motion. In Task and Motion Planning (TAMP), a high-level action like PickUp(cup) is only valid if there exists a kinematically feasible, collision-free trajectory to execute it.
When human experts define these domains manually, they must anticipate every physical constraint. Brandon Vu et al. observe that the manual design of these domain-specific constraints limits the portability of planners across different robot platforms and environments. Furthermore, in dynamic or uncertain settings, writing down deterministic rules becomes impractical. Sergio Jiménez et al. emphasize that manually defining all possible action outcomes and their probabilities in stochastic environments is highly error-prone. Early work by Xuemei Wang in Learning by Observation and Practice: A Framework for Automatic Acquisition of Planning Operators recognized that manually specifying preconditions and effects restricts scalability, proposing instead that agents should refine these operators through iterative practice.
Inventing Symbols from Continuous Data
How can a robot autonomously discover its own symbolic vocabulary from raw, continuous sensorimotor data? The core intuition is that a "good" symbolic predicate is one that makes high-level planning both feasible and computationally efficient.
Tom Silver et al. formalized this intuition in Predicate Invention for Bilevel Planning. Instead of asking humans to invent concepts like Holding(block), their framework automatically invents predicates by optimizing a surrogate objective that directly estimates planning time from demonstrations. If a newly proposed predicate helps partition the continuous state space such that a classical planner can find a solution faster, the predicate is retained. This bridges the gap between continuous geometry and discrete logic without human intervention.
Similarly, Leon Keller et al. propose neuro-symbolic imitation learning to discover symbols and skills directly from demonstrations, allowing robots to build their own structured "language" for multi-step tasks. By learning these abstractions from data, the agent bypasses the need for the predefined sorted signatures of objects and actions that Mohan Sridharan notes typically constrain traditional hybrid architectures.
Inferring Domains from Demonstrations and Feedback
Another powerful approach to bypassing the manual modeling bottleneck is to infer the underlying PDDL structure from minimal human input, such as a few video demonstrations or natural language corrections.
In Automated Planning Domain Inference for Task and Motion Planning, Jinbang Huang et al. demonstrate how a system can infer complete TAMP domains from just a handful of human demonstrations. They utilize a Graph Attention Network (GAT) to predict relevant logical components combined with a generate-and-test search algorithm, producing minimal, effective PDDL domains that outperform standard behavior cloning.
When demonstrations are ambiguous, interactive feedback can fill the gaps. Muzhi Han et al. introduced the InterPreT framework, which allows non-expert humans to provide natural language corrections when a robot fails a task. The system uses a Large Language Model (LLM) to translate this interactive feedback into executable Python code and formal PDDL domains, enabling the robot to learn the symbolic language of task planning on the fly. This shifts the burden of knowledge engineering from specialized AI experts to intuitive, interactive human-robot communication.
Translating Vision and Language to Formal Logic
The rise of foundation models has provided a new mechanism for overcoming the domain modeling bottleneck: using the implicit world knowledge embedded in Large Language Models (LLMs) and Vision-Language Models (VLMs) to generate formal planning representations.
Keisuke Shirai et al. developed ViLaIn (Vision-Language Interpreter) to automatically translate raw images and natural language instructions into machine-readable PDDL files. This eliminates the need for humans to write code-like problem descriptions for every new scenario. To handle physical realities, Jeremy Siburian et al. extended this concept into ViLaIn-TAMP, introducing a corrective planning loop. When a low-level motion planner fails due to a geometric constraint, the detailed physical failure feedback is fed back into the VLM, which dynamically updates the symbolic domain model.
By grounding foundation models within the rigorous constraints of classical planners, these hybrid systems eliminate the labor-intensive process of manual domain specification while retaining the mathematical guarantees of traditional AI planning.