GET-Zero: Graph Embodiment Transformer for Zero-shot Embodiment Generalization

Abstract

This paper introduces GET-Zero, a model architecture and training procedure for learning an embodiment-aware control policy that can immediately adapt to new hardware changes without retraining. To do so, we present Graph Embodiment Transformer (GET), a transformer model that leverages the embodiment graph connectivity as a learned structural bias in the attention mechanism. We use behavior cloning to distill demonstration data from embodiment-specific expert policies into an embodiment-aware GET model that conditions on the hardware configuration of the robot to make control decisions. We conduct a case study on a dexterous in-hand object rotation task using different configurations of a four-fingered robot hand with joints removed and with link length extensions. Using the GET model along with a self-modeling loss enables GET-Zero to zero-shot generalize to unseen variation in graph structure and link length, yielding a 20% improvement over baseline methods. All code and qualitative video results are on https://get-zero-paper.github.io

Figures (5)

• Figure 1: GET-Zero is an embodiment-aware policy that is able to zero-shot generalize to unseen embodiment designs with varied geometry, number of joints, and graph structure.
• Figure 2: Graph Embodiment Transformer (GET). GET is an embodiment-aware model based on a transformer encoder. Each joint forms separate tokens containing local sensory and embodiment information. The self-attention layers use an undirected (Spatial Bias) and directed (Parent-Child Bias) graph distance to bias the attention scores between joints according to the embodiment graph (grid color intensity indicates distance between nodes). A policy head predicts actions and a self modeling head predicts meta-properties about the embodiment, such as forward kinematics.
• Figure 3: Training procedure. To train GET-Zero, we follow a teacher-student paradigam, where the teachers are separate embodiment-specific experts trained using RL. Then we distill knowledge from the experts into a single embodiment-aware transformer (i.e., the student policy) using behavior cloning with a self-modeling loss. GET-Zero takes as input embodiment definition and proprioception, and infers proper actions to perform an in-hand rotation task for an unseen embodiment.
• Figure 4: Finger Variations. We procedurally generate variations of the fingers in the LEAP Hand leaphand by removing joints and links as well as adding in 1.5cm link length extensions (orange).
• Figure 5: Impact of training embodiments on zero-shot graph generalization. We observe that even with fewer training embodiments, GET-Zero achieves reasonable in-hand rotation performance (5 seeds).