[Paper Review] Unsupervised Emergence of Egocentric Spatial Structure from Sensorimotor Prediction

Alban Laflaquière, M. Ortíz|City Research Online (City University London)|Jun 4, 2019

Action Observation and Synchronization34 references4 citations

TL;DR

This paper proposes an unsupervised sensorimotor prediction framework that enables an agent to autonomously learn a topologically and metrically accurate egocentric spatial representation from raw sensorimotor experience, without prior knowledge or supervision. The key result is that a neural network trained to predict future sensory states captures the underlying Euclidean structure of space as a byproduct of learning sensorimotor invariants.

ABSTRACT

Despite its omnipresence in robotics application, the nature of spatial knowledge and the mechanisms that underlie its emergence in autonomous agents are still poorly understood. Recent theoretical works suggest that the Euclidean structure of space induces invariants in an agent's raw sensorimotor experience. We hypothesize that capturing these invariants is beneficial for sensorimotor prediction and that, under certain exploratory conditions, a motor representation capturing the structure of the external space should emerge as a byproduct of learning to predict future sensory experiences. We propose a simple sensorimotor predictive scheme, apply it to different agents and types of exploration, and evaluate the pertinence of these hypotheses. We show that a naive agent can capture the topology and metric regularity of its sensor's position in an egocentric spatial frame without any a priori knowledge, nor extraneous supervision.

Motivation & Objective

• To investigate whether an unsupervised agent can learn a structured egocentric spatial representation from sensorimotor experience alone.
• To test if sensorimotor prediction can capture topological and metric invariants of external space.
• To evaluate whether such spatial structure emerges without prior knowledge, hand-designed inductive biases, or extraneous supervision.
• To examine the robustness of spatial representation learning across different agents, exploration strategies, and environmental complexities.

Proposed method

• A deep neural network is trained to predict future sensory states from current sensory input and motor actions.
• The network uses a shared encoder (Net_enc) to map sensory and motor inputs into a joint latent representation h.
• The model is trained end-to-end via supervised prediction loss on future sensory states, using a predictive autoencoder setup.
• Three exploration strategies—random, MEM (memory-efficient motion), and regular motor sampling—are used to evaluate generalization.
• The latent representation h is evaluated for spatial structure using topological (D_topo) and metric (D_metric) dissimilarity measures against ground-truth sensor positions.
• The framework is tested on two environments: a discrete 2D grid world and a continuous 3D robotic arm simulation with 6 degrees of freedom.

Experimental results

Research questions

• RQ1 Can an agent learn a topologically accurate egocentric spatial representation through unsupervised sensorimotor prediction?
• RQ2 Does the learned representation capture the metric regularity of space without explicit supervision or hand-designed priors?
• RQ3 How robust is the emergence of spatial structure to variations in agent morphology, exploration strategy, and environmental complexity?
• RQ4 To what extent do sensorimotor invariants—particularly compensable changes—give rise to a Euclidean spatial concept in the latent space?
• RQ5 Is the spatial structure in the latent representation independent of environmental content and sensory modality?

Key findings

• The latent representation h successfully captures the topological structure of the agent’s sensor position, as evidenced by low D_topo values across all environments and exploration types.
• The metric regularity of space is preserved in h, with D_metric values indicating that distances in the latent space correlate with true spatial distances in the environment.
• The emergence of spatial structure is robust to increased agent complexity, as shown in the 6-DOF arm setup where similar spatial regularity was observed despite higher-dimensional motor spaces.
• The results remain consistent across different exploration strategies, including random and MEM exploration, indicating generalization beyond specific motor patterns.
• The use of a log scale in visualization was necessary due to higher dissimilarity values in MEM exploration, but the underlying spatial structure remained preserved.
• Training variability was found to be primarily due to stochasticity in initialization and mini-batch selection, not dataset differences, as shown by consistent standard deviations across 50 independent runs.