Quasimetric Embeddings for Non-Reversible Environments

170 concepts76 papers1 proposal

In the idealized world of many reinforcement learning (RL) benchmarks, moving from point A to point B is exactly as "costly" as moving from B back to A. However, the physical world is rarely so cooperative. For a robot, climbing a steep hill consumes far more energy than coasting down it. Some actions, like driving off a ledge or breaking a glass, are entirely irreversible.

Standard machine learning models often rely on symmetric distance metrics (like Euclidean distance) to represent the relationship between states. These metrics assume $d(A, B) = d(B, A)$. When this assumption meets a non-reversible environment, the model's "map" of the world breaks, leading to sub-optimal planning and unsafe behavior. Recent research has converged on a solution: quasimetric embeddings, a mathematical framework that respects the "arrow of time" and the directional costs of the real world.

The Geometry of One-Way Streets

A quasimetric is a distance function that satisfies the requirements of a standard metric—it is non-negative and obeys the triangle inequality—but it discards the requirement of symmetry. In a quasimetric space, the distance from $A$ to $B$ can be vastly different from the distance from $B$ to $A$.

As Anders Jonsson et al. argue, enforcing the triangle inequality ($d(x, z) \le d(x, y) + d(y, z)$) is the critical "glue" that allows for multi-step planning. It ensures that the learned representation remains mathematically coherent; if an agent knows the cost to get from $A$ to $B$ and from $B$ to $C$, the triangle inequality provides a reliable upper bound for the total journey. By using asymmetric "Wide Norms," the authors demonstrate that agents can navigate environments with one-way paths or varying terrain difficulty far more effectively than those using symmetric metrics, which suffer from an "averaging effect" that blurs easy and difficult paths.

Decomposing Potential and Residual Costs

Why does asymmetry work so well for physical systems? One intuition is that traversal costs can be split into two parts: a reversible potential (like distance or elevation change) and a dissipative residual (like energy lost to friction or gravity).

Bo Liu et al. proposed the Metric Residual Network (MRN), which explicitly bakes this decomposition into a neural architecture. The symmetric component provides a fast general approximation of distance, while the asymmetric residual accounts for directional constraints like gravity. Similarly, Jumman Hossain et al. introduce Quasi-Potential Functions, which separate state-dependent potential from path-dependent dissipative costs. This structural split allows robots to navigate rugged terrain up to 13.6% more energy-efficiently by recognizing that "uphill" and "downhill" are fundamentally different tasks [Jumman Hossain et al.].

Capturing the Arrow of Time

In sequential decision-making, asymmetry is not just about energy—it is about reachability. Time only flows forward, and many state transitions are one-way. Nasim Rahaman et al. introduced the $h$-potential, a scalar function that learns the "arrow of time" by quantifying the inherent asymmetry of state transitions. This allows agents to distinguish between states that are easily accessible and those that are "downstream" and impossible to return from.

This temporal asymmetry is a recurring theme across the field:

  • Foundation Policies: Seohong Park et al. show that representing environment reachability with symmetric metrics fails in non-reversible Markov Decision Processes, such as jumping down a ledge.
  • State Chrono Representation: Jianda Chen et al. use quasimetric embeddings to model the one-way progression of trajectories, creating features that are more robust to visual distractions.
  • Unsupervised Learning: Junik Bae utilizes "temporal distance"—the number of steps between states—as a directional compass for robots to explore complex environments without human rewards.

Beyond Robotics: Asymmetry in Vision and Preferences

The utility of asymmetric modeling extends beyond physical navigation into abstract feature spaces. In computer vision, Wenhao Wang et al. point out that image relationships are often directional: a small crop is a derivative of a larger original image, but the original is not a derivative of the crop. By using asymmetric similarity modeling, they can better detect image copies.

In the realm of human behavior, Adish Singla et al. use hemimetrics (another term for quasimetrics) to model "switching costs." They observe that the psychological cost of downgrading from a high-quality service to a lower one is rarely the same as the cost of upgrading.

Finally, in deep metric learning, Hatice Kubra Cilingir et al. demonstrate that Deep Bregman Divergences—which are inherently asymmetric—outperform standard Euclidean distances in classification tasks. This suggests that directional relationships in feature space capture more nuanced information than simple reciprocal similarity.

Why It Works: The Power of Directed Constraints

The core intuition behind this research direction is that constraints are information. By forcing a model to acknowledge that $A \to B$ is not the same as $B \to A$, we provide it with a more accurate prior about the structure of the world.

As Sven Koenig noted in early work on agent-centered search, considering potential irreversibility ensures that local decisions don't lead to high-cost recovery paths. Whether it is a robot avoiding a point of no return [Jeffrey Johns et al.], a vehicle routing system navigating one-way streets [RADAR], or a curriculum learning strategy focusing on the "frontier" of an agent's capabilities [Jan Wöhlke], asymmetric distance modeling allows AI to navigate a world that is fundamentally directed.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 79 papers
  2. Version 1 · 2026-10-04 · 118 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_asymmetric_distance_modeling_9af4ea12,
  title = {asymmetric distance modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/asymmetric-distance-modeling/rdr_9af4ea12fcc9420c8ca92060502ca58d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_asymmetric_distance_modeling_9af4ea12_v2,
  title = {asymmetric distance modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/asymmetric-distance-modeling/rdr_9af4ea12fcc9420c8ca92060502ca58d/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake