Imagine That! Leveraging Emergent Affordances for 3D Tool Synthesis
In this paper we explore the richness of information captured by the latent space of a vision-based generative model – and how to exploit it. The context of our work is an artificial agent’s ability to perform task-focused tool synthesis for 3D reaching tasks based purely on 2D visual inputs. In particular, given visual observations of a reaching task and a proposed tool, our approach employs activation maximisation of a task-based performance predictor to directly optimise the 3D geometry of the tool by traversing a learnt latent space. While the embedding learned by the generative model captures the factors of variation in 3D tool geometry, e.g. length, width and configuration, the performance predictor identifies sub-manifolds correlated with task success in a weakly supervised manner. Using a 3D simulation environment, we demonstrate that traversing the latent space in this task-driven way results in tool geometries appropriate for the task at hand.Our results therefore suggest that affordances – like the utility for reaching – are encoded along smooth trajectories in the learned latent space. Accessing these emergent affordances via gradient descent considering only high-level performance criteria (such as task success) enables the agent to manipulate tool geometries in a targeted and deliberate way.