Bilevel Planning with Learned Symbolic Abstractions from Interaction Data

Fatih Doğangün, Burcu Kilic, Serdar Bahar, Emre Uğur

TL;DR

Researchers have developed a bilevel neuro-symbolic framework that allows robots to plan complex manipulation tasks by combining the speed of abstract symbolic reasoning with the precision of continuous dynamics. By using a dual-model approach—one to "invent" symbols and another to verify them—the system achieves the high success rates of exhaustive continuous search while solving most problems through much faster symbolic logic.

What Problem Does This Paper Address?

Robots operating in the real world face a "representation gap." Planning directly in high-dimensional, continuous sensorimotor space (e.g., exact 3D coordinates) is computationally exhausting. Conversely, traditional AI planning uses discrete symbols (e.g., "Block A is on Block B"), which are fast but "lossy"—they often ignore the messy geometric realities that cause a plan to fail in practice.

While recent "Deep Symbolic" methods can learn these abstractions from raw data, they typically rely on deterministic rules and lack a way to check if an abstract plan will actually work in the real world. This paper addresses the need for a planning framework that is both efficient (using symbols) and robust (verifying against continuous physics).

Methods

The authors propose a Bilevel Planning Framework supported by a Dual-Model Scheme:

  1. The Symbolic Model ($M_{sym}$): This neural network uses a Gumbel-Sigmoid bottleneck to force continuous data into discrete, relational symbols (like "object $i$ is near object $j$"). It learns probabilistic rules (PPDDL) from robot interaction data.
  2. The Dynamics Model ($M_{dyn}$): A twin network that uses standard Sigmoid activations. It is optimized for high-precision continuous effect prediction rather than discretization.

The Planning Pipeline (Algorithm 1): * Level 1 (Symbolic): The robot generates multiple candidate plans using the probabilistic symbolic rules. It then uses the Dynamics Model to "mentalize" or simulate these plans in continuous space. If a plan's predicted final state is within a tolerance threshold ($\tau_{verify}$), it is executed. * Level 2 (Fallback): If no symbolic plan passes verification, the system falls back to a more expensive Continuous Forward Search (Weighted A*) using the Dynamics Model to find a path from scratch.

Datasets

The models were trained and evaluated using:

  • Interaction Data: A dataset of 140,000 interaction samples collected from a simulated tabletop environment.
  • Environment: A UR10e manipulator interacting with varying numbers (2, 3, or 4) of large and small blocks.
  • Evaluation Set: 100 unique planning problems for each combination of object and action counts (up to 5-step plans).

Key Insights and Ideas

  • The Specialization Paradox: The authors found that a single model cannot be "jack of all trades." A model forced to discretize (for symbols) loses the precision needed for physics prediction. By splitting the architecture into a "Symbolic Specialist" and a "Dynamics Specialist," they achieved better performance in both domains.
  • Probabilistic > Deterministic: Traditional symbolic planners often fail because they only consider the most likely outcome of an action. By sampling an ensemble of possible symbolic domains, the robot can discover "less likely" but successful paths that a deterministic planner would ignore.
  • Verification as a Bridge: The continuous dynamics model acts as a "sanity check." This prevents the robot from attempting symbolic plans that are logically sound but physically impossible due to accumulated errors or geometric constraints.

Key Results

  • High Reliability: The verification mechanism achieved a mean accuracy of 0.91 and an F1 score of 0.90 in identifying which symbolic plans would succeed or fail.
  • Efficiency vs. Power: The Bilevel Planner achieved success rates statistically comparable to pure continuous search (the gold standard for accuracy) but resolved the majority of problems using efficient symbolic reasoning.
  • Superiority over Baselines: The probabilistic approach significantly outperformed the deterministic baseline (from prior work), especially as the complexity of the task (number of actions) increased. For 3-action tasks, the probabilistic planner improved success rates from ~36% to ~45% before even applying the bilevel fallback.

Limitations and Future Work

  • Data Scarcity for Rare Events: The current models rely on random exploration. This means they rarely see "critical" or rare state transitions, which limits the complexity of the rules they can learn. The authors suggest active learning or curriculum learning to solve this.
  • Fixed Thresholds: The verification threshold ($\tau_{verify}$) is currently a fixed value (7cm). In the future, an adaptive threshold that adjusts based on the number of objects or the specific task could improve the "Level 1 to Level 2" transition.
  • Heuristic Pruning: While the continuous search is effective, it still requires heavy pruning of the action space to remain tractable. Integrating more sophisticated reachability heuristics could further speed up the fallback mechanism.

Summary updated

Abstract

Intelligent agents must reason over both continuous dynamics and discrete representations to generate effective plans in complex environments. Previous studies have shown that symbolic abstractions can emerge from neural effect predictors trained with a robot's unsupervised exploration. However, these methods rely on deterministic symbolic domains, lack mechanisms to verify the generated symbolic plans, and operate only at the abstract level, often failing to capture the continuous dynamics of the environment. To overcome these limitations, we propose a bilevel neuro-symbolic framework in which learned probabilistic symbolic rules generate candidate plans rapidly at the high level, and learned continuous effect models verify these plans and perform forward search when necessary at the low level. Our experiments on multi-object manipulation tasks demonstrate that the proposed bilevel method outperforms symbolic-only approaches, reliably identifying failing plans through verification, and achieves planning performance statistically comparable to continuous forward search while resolving most problems via efficient symbolic reasoning.
BibTeX
@misc{dogangun2026bilevel,
  author = {Fatih Doğangün and Burcu Kilic and Serdar Bahar and Emre Uğur},
  title = {{Bilevel Planning with Learned Symbolic Abstractions from Interaction Data}},
  year = {2026},
  eprint = {2603.08599},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.08599},
}
Figure