Relative Coordinate Frames for Generalizable Control

146 concepts113 papers2 proposals

When training a reinforcement learning agent or designing a robotic controller, the choice of coordinate system is often treated as a minor implementation detail. However, relying on a fixed, global coordinate system forces a model to learn the absolute position of every object in the workspace. If a robot learns to pick up a cup at coordinates $(x_1, y_1, z_1)$, that knowledge does not inherently transfer if the cup is moved to $(x_2, y_2, z_2)$, even if the physical mechanics of the grasp are identical.

To build policies that generalize across environments, scales, and hardware configurations, a prominent research direction focuses on relative state representations. By transforming absolute spatial data into local, task-relevant coordinate frames, models can exploit geometric symmetries and focus entirely on the relative dynamics of an interaction.

The Core Intuition: Why Relative Frames Work

Relative state representations fundamentally alter what an agent "sees" by anchoring the origin of the coordinate system to a dynamic entity—such as the agent itself, a target object, or a moving goal. This shift provides several mathematical and practical advantages.

1. Dimensionality Reduction and State Space Bounding

In a global coordinate system, the state space expands with the size of the environment. An agent must track its own position and the positions of all other entities relative to a distant, arbitrary origin. By switching to an egocentric or agent-centric frame, the agent's own position and orientation become constants (the origin).

As demonstrated by Park et al. in Kick-motion Training with DQN in AI Soccer Environment, transitioning from absolute coordinates to an egocentric frame can reduce state space dimensionality by up to 60%. This reduction eliminates redundant global variables, allowing learning algorithms to converge significantly faster by focusing solely on the local geometric relationships required for immediate action.

2. Spatial Invariance and Zero-Shot Transfer

The primary strength of relative representations is their ability to bake translation and rotation invariance directly into the model's inputs. If a task is represented purely through relative distances, angles, and velocities, the learned policy becomes invariant to where the task occurs in the world.

For instance, Kofinas et al. in Roto-translated Local Coordinate Frames For Interacting Dynamical Systems introduced local coordinate frames per object to grant graph neural networks Galilean invariance, allowing models to forecast trajectories without being distracted by arbitrary global positioning. Similarly, Chiu et al. show in Bimanual Regrasping for Suture Needles using Reinforcement Learning for Rapid Motion Planning that defining a state space using relative poses between end-effectors and target objects yields robot-agnostic policies that generalize across different hardware configurations and base placements.

This spatial invariance enables zero-shot transfer. In Zero-Shot Transfer with Deictic Object-Oriented Representation in Reinforcement Learning, Marom et al. utilize "deictic" (relative) representations centered on a focal object. Because the interaction rules remain invariant regardless of environment size or object count, agents can master a task in a small environment and immediately apply it to a much larger one.

3. Mitigating Non-Stationarity and Environmental Drift

Neural networks struggle with inputs that grow indefinitely or drift over time. In long-horizon tasks or continuous navigation, absolute coordinates are non-stationary variables that can easily push a model into out-of-distribution (OOD) state values.

Kim et al. highlight this in Observation Space Matters: Benchmark and Optimization Algorithm, showing that removing absolute global coordinates or converting them to relative coordinates prevents the agent from failing when encountering unseen spatial ranges. In physical laboratory settings, Richtmann et al. demonstrate in Model-free reinforcement learning with noisy actions for automated experimental control in optics that excluding absolute positions forces a controller to rely on relative signal changes, making the system highly resilient to physical environmental drift over time.

A Taxonomy of Relative Reference Frames

Depending on the task, researchers anchor relative coordinate transformations to different focal points:

Agent-Centric and Egocentric Frames

Here, the coordinate system moves with the agent. In Obstacle Avoidance for UAS in Continuous Action Space Using Deep Reinforcement Learning, Hu et al. transform observations so the agent remains at the origin with the goal aligned along a fixed axis, allowing the model to learn universal geometric principles of obstacle avoidance. Johnson et al. apply a similar concept in Dynamically Constrained Motion Planning Networks for Non-Holonomic Robots using an egocentric costmap to normalize environmental structures relative to a robot's current position.

Object and Part-Centric Frames

When the task involves manipulating specific items, anchoring the frame to the object or its sub-parts isolates the interaction mechanics from the rest of the world. Geng et al. use part-based canonicalization in PartManip: Learning Cross-Category Generalizable Part Manipulation Policy from Point Cloud Observations to standardize state information relative to a target object part. This ensures the policy remains robust to variations in the object's absolute position within a workspace.

Goal and Setpoint-Centric Frames

Instead of tracking where the robot is, the state is framed around how far the robot is from its intended target. This idea dates back to early control scheduling work; Guha in A Reinforcement Learning Variant for Control Scheduling encoded system states relative to a moving target setpoint rather than absolute coordinates, enabling a single learned policy to generalize across diverse operating ranges. More recently, Lee showed in Low-level Pose Control of Tilting Multirotor for Wall Perching Tasks Using Reinforcement Learning that expressing a robot's state relative to its target goal reduces the search space and accelerates policy convergence.

Multi-Agent and Relational Abstractions

In multi-agent systems, absolute coordinates scale poorly as the number of agents increases. Researchers instead use relative configurations or relational predicates. Steleac et al. show in Inter-Agent Relative Representations for Multi-Agent Option Discovery that centering multi-agent states on a dynamic reference point (like the team's geometric center of gravity) ensures coordination strategies remain invariant to absolute spatial positions.

For formal safety verification, Omidshafiei et al. introduce a relative coordinate frame in Collision Avoidance Verification of Multiagent Systems with Learned Policies to create a stationary representation of collision sets, simplifying multi-agent verification into a pair-wise single-agent problem relative to a fixed reference point.

Conclusion

By shifting the analytical perspective from "Where am I in the world?" to "Where am I relative to what matters?", relative state representations bypass the need for models to memorize spatial layouts. Whether through egocentric costmaps, deictic symbols, or object-centric canonicalization, this representational strategy provides a mathematically rigorous foundation for sample-efficient, robust, and generalizable control.

Key Foundational Work

A Reinforcement Learning Variant for Control Scheduling (1990) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_relative_state_representation_258905bc,
  title = {relative state representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/relative-state-representation/rdr_258905bc86d14f68870e370eafb6d8c5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_relative_state_representation_258905bc_v1,
  title = {relative state representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/relative-state-representation/rdr_258905bc86d14f68870e370eafb6d8c5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake