ALAN: Autonomously Exploring Robotic Agents in the Real World

TL;DR

ALAN tackles scalable real-world robotic learning by coupling an environment-change signal with agent-centric uncertainty to drive autonomous exploration, using a world model (RSSM) and planning via Cross-Entropy Method. It also leverages Detic-based visual priors and a kNN-driven goal-reaching mechanism to achieve zero-shot manipulation goals from goal images, demonstrated on two real play kitchens with a Franka Panda in under 150 trajectories. The key contributions are the dual-signal exploration framework, the integration of offline visual priors to focus data collection, and empirical validation showing improved exploration data quality and robust goal attainment in real-world manipulation. This work advances practical autonomous robotics by reducing reliance on task-specific rewards and human demonstrations while enabling efficient data collection and transfer to user-defined goals.

Abstract

Robotic agents that operate autonomously in the real world need to continuously explore their environment and learn from the data collected, with minimal human supervision. While it is possible to build agents that can learn in such a manner without supervision, current methods struggle to scale to the real world. Thus, we propose ALAN, an autonomously exploring robotic agent, that can perform tasks in the real world with little training and interaction time. This is enabled by measuring environment change, which reflects object movement and ignores changes in the robot position. We use this metric directly as an environment-centric signal, and also maximize the uncertainty of predicted environment change, which provides agent-centric exploration signal. We evaluate our approach on two different real-world play kitchen settings, enabling a robot to efficiently explore and discover manipulation skills, and perform tasks specified via goal images. Website at https://robo-explorer.github.io/

Figures (7)

• Figure 1: We present ALAN, an approach for real world robotic exploration in challenging manipulation environments.
• Figure 2: We propose Autonomous Learning Agents (ALAN) that can enable robots to collect rich data from their environment efficiently. The agent utilizes environment change, both directly as an environment-centric signal, as well as modelling the change and taking actions that maximize uncertainty in change space, which provides agent-centric signal.
• Figure 3: Visualizations of the object detections, using detic. The masks selected to study exploration are the knife, pan and rightcabinet handle from kitchen1 (left), and the topshelf, fridge handles and pot from kitchen2 (right).
• Figure 4: An example of the change image extracted from a pair of images, as described in Equation \ref{['eq:ec_rew']}. This is a binary image that detects pixels where change has occured.
• Figure 5: We explore on 6 settings across two play kitchens. Top, from left: cabinet, knife, pan (kitchen1). Bottom, from left: top shelf, pot, fridge (kitchen2).