[Paper Review] Imagine That! Leveraging Emergent Affordances for 3D Tool Synthesis

Yizhe Wu, Sudhanshu Kasewa+5 more|arXiv (Cornell University)|Sep 30, 2019

Manufacturing Process and Optimization47 references4 citations

TL;DR

This paper proposes a method for 3D tool synthesis by leveraging emergent affordances in the latent space of a vision-based generative model. By using a task-based performance predictor to guide activation maximization in the latent space, the model imagines and generates task-appropriate tools through smooth, interpretable trajectories in geometry (e.g., length, width, shape), achieving 83.8% success on unseen reaching tasks—demonstrating that affordances emerge as structured paths in latent space.

ABSTRACT

In this paper we explore the richness of information captured by the latent space of a vision-based generative model. The model combines unsupervised generative learning with a task-based performance predictor to learn and to exploit task-relevant object affordances given visual observations from a reaching task, involving a scenario and a stick-like tool. While the learned embedding of the generative model captures factors of variation in 3D tool geometry (e.g. length, width, and shape), the performance predictor identifies sub-manifolds of the embedding that correlate with task success. Within a variety of scenarios, we demonstrate that traversing the latent space via backpropagation from the performance predictor allows us to imagine tools appropriate for the task at hand. Our results indicate that affordances-like the utility for reaching-are encoded along smooth trajectories in latent space. Accessing these emergent affordances by considering only high-level performance criteria (such as task success) enables an agent to manipulate tool geometries in a targeted and deliberate way.

Motivation & Objective

• To investigate whether task-driven exploration of a structured latent space can generate meaningful, task-appropriate 3D tools.
• To explore whether object affordances—such as reachability—emerge as smooth trajectories in the latent space of a generative model.
• To develop a hybrid architecture combining unsupervised generative learning with a high-level performance predictor for targeted tool manipulation.
• To demonstrate that high-level task success criteria can guide interpretable geometric modifications in 3D tool shapes without explicit supervision on shape factors.

Proposed method

• A variational autoencoder (VAE) is trained on 3D tool meshes to learn a structured latent space capturing geometric factors like length, width, and shape.
• A task-based performance predictor (a classifier) is trained to predict success in a reaching task using visual observations of the scenario and tool geometry.
• During inference, activation maximization is applied to the performance predictor's output to guide backpropagation through the latent space, modifying the tool's latent representation toward higher predicted success.
• The updated latent code is decoded into a new 3D mesh, enabling the agent to 'imagine' improved tool geometries for the given task.
• The method is evaluated against a task-unaware model and a random walk baseline to isolate the contribution of task-guided optimization.
• The model is trained end-to-end with a joint objective: the VAE reconstruction loss and the performance predictor's cross-entropy loss, shaping the latent space to encode task-relevant disentangled factors.

Experimental results

Research questions

• RQ1 Can emergent affordances such as reachability be encoded as smooth trajectories in the latent space of a generative model?
• RQ2 Can a high-level performance predictor guide meaningful, interpretable modifications in 3D tool geometry through latent space optimization?
• RQ3 Does task-driven traversal of the latent space lead to more effective tool synthesis than random or task-agnostic exploration?
• RQ4 To what extent can the model generalize to unseen scenario types during tool imagination?

Key findings

• The task-driven model achieved a global success rate of 83.8% in generating tools for unseen reaching tasks, significantly outperforming the task-unaware baseline (63.6%).
• The random walk baseline, which explores the latent space stochastically, failed to find suitable tools, indicating that unguided exploration is ineffective.
• Qualitative results show that the model generates novel tool shapes—such as T-shaped or hook-like tools—by interpolating between distinct geometric components like vertical and horizontal parts.
• The model learns to modify interpretable geometric factors such as length, width, and configuration in a smooth, continuous manner, suggesting that these factors are encoded as trajectories in the latent space.
• The performance predictor successfully identifies and guides the model along sub-manifolds in the latent space that correspond to task-relevant affordances, such as improved reachability.
• The results demonstrate that affordances emerge naturally as structured paths in the latent space when shaped by a high-level task objective, rather than being explicitly encoded.