Task Parameter Extrapolation via Learning Inverse Tasks from Forward Demonstrations

TL;DR

This work constructs a common representation of the forward and inverse tasks, and leverages auxiliary forward demonstrations from novel configurations to successfully execute the corresponding inverse tasks, without any direct supervision.

Abstract

Generalizing skill policies to novel conditions remains a key challenge in robot learning. Imitation learning methods, while data-efficient, are largely confined to the training region and consistently fail on input data outside it, leading to unpredictable policy failures. Alternatively, transfer learning approaches offer methods for trajectory generation robust to both changes in environment or tasks, but they remain data-hungry and lack accuracy in zero-shot generalization. We address these challenges by framing the problem in the context of task inversion learning and proposing a novel joint learning approach to achieve accurate and efficient knowledge transfer. Our method constructs a common representation of the forward and inverse tasks, and leverages auxiliary forward demonstrations from novel configurations to successfully execute the corresponding inverse tasks, without any direct supervision. We show the extrapolation capabilities of our framework via ablation studies and experiments in simulated and real-world environments that require complex manipulation skills with a diverse set of objects and tools, where we outperform diffusion-based alternatives.

Figures (8)

• Figure 1: The demonstrations are decoupled into task and sensorimotor (SM) spaces. In the top, demonstrations are visualized as lines between task parameters ($\psi$) and SM executions ($\tau$). Our method will first identify correspondence between demonstrations and learns a common representation of the forward and inverse tasks using paired and auxiliary datasets. The learning objective is to infer inverse executions for task parameters drawn from the auxiliary distribution.
• Figure 2: Overview of the proposed method. The first and second rows show the overall paired and auxiliary passes, respectively, during training. The third row shows the inference conditioned on the novel task parameter and observations from a forward execution to generate the full inverse execution.
• Figure 3: Visualization of the synthetic trajectory datasets. The color correspondence illustrates the pairing. The trajectories for the forward task is on the upper left. Remain plots show the corresponding inverse task trajectories under different conditions: randomly paired (Random), paired using our algorithm with noisy task parameters (Noisy), and perfectly paired (Perfect).
• Figure 4: Task controller logic. The flowchart illustrates the underlying decision process in provided demonstrations.
• Figure 5: Visualization of the manipulation tasks in the simulated environment. The figure shows example executions for the designated forward task (moving an object to a target) and its inverse task with three different modalities for various object categories (left to right, top to bottom: large horizontal cylinder, sphere, vertical cylinder, large box, small horizontal cylinder, small box).