Abstract
This paper introduces a low-cost, end-to-end imitation learning framework designed to make training industrial-grade robots accessible to individual researchers. By combining off-the-shelf hardware, such as VR controllers, with a Diffusion Policy-based architecture, the authors demonstrate that multi-task robot learning can be achieved with a relatively small number of demonstrations. The study also contributes a dataset of over 4,000 episodes across 10 tasks and proposes the Voting Positive Rate (VPR), a new evaluation metric to standardize real-world performance testing.
Introduction
Imitation learning holds significant promise for teaching robots complex manipulation tasks, yet the barrier to entry remains high due to the need for expensive collaborative robots (cobots), specialized laboratory setups, and massive datasets. Furthermore, evaluating real-world performance is often subjective, leading to inconsistencies in reporting success rates. The Generalized Robot Learning Framework (GRLF) addresses these challenges by providing an affordable, hardware-agnostic pipeline and a rigorous human-evaluation metric to minimize bias.
Methodology
Hardware and Data Collection
The authors developed a cost-effective data collection setup utilizing an industrial robotic arm, two standard RGB-D cameras (one wrist-mounted and one global), and an Oculus Quest 2 controller for intuitive teleoperation. This setup allows for high-quality data collection without the need for expensive proprietary haptic devices.
Software Architecture
The system employs a Diffusion Policy approach, bifurcated into two primary components:
- Perception Module: This module encodes visual and state data. The researchers evaluated standard CNNs (ResNet18/34) and Feature Pyramid Networks (FPN).
- Action Prediction Module: Utilizing Denoising Diffusion Probabilistic Models (DDPM), this module generates robot trajectories. The study compared CNN-based (UNet) and Transformer-based architectures to determine their efficacy in handling temporal dependencies.
Standardized Evaluation: Voting Positive Rate (VPR)
To mitigate the subjectivity of real-world testing, the authors introduced the Voting Positive Rate (VPR). In this protocol, four independent evaluators judge a robot's performance across various initial conditions. A trial is classified as a success only if all four evaluators vote positively, ensuring a rigorous benchmark suitable for industrial standards.
Datasets and Tasks
The researchers released a custom open-source dataset containing over 4,000 episodes across 10 distinct robotic tasks. These tasks range from basic pick-and-place operations to complex, multi-step logic challenges, such as sorting objects by size, stacking specific cups, and ring tossing. The dataset is hosted on HuggingFace to facilitate further research in low-cost imitation learning.
Key Insights and Results
- Architectural Performance: The combination of an FPN-based ResNet34 perception module and a Transformer-based action module consistently outperformed other configurations, particularly on tasks requiring sequential logic. Transformers proved superior at capturing long-range spatial and temporal dependencies compared to CNN-based UNet models.
- Data Efficiency: Increasing the number of human demonstrations significantly improved performance. For example, doubling demonstrations from 60 to 120 on the "PickSmall" task resulted in a 10% increase in success rate. On the complex "PickBigV2" task, the FPN/Transformer model achieved a 78.3% VPR with 240 demonstrations.
- Transfer Learning: Multi-task pre-training proved highly efficient; a model pre-trained on a "Basketball" task could be fine-tuned for a "BlockPick" task in just 50 additional epochs, compared to the hundreds required for training from scratch.
- Generalization: The model demonstrated zero-shot generalization to unseen camera angles during deployment after being trained on only two distinct global views.
Conclusion and Future Work
While the GRLF successfully lowers the entry barrier for robot learning, it remains dependent on the quality and quantity of human demonstrations. Future research will focus on leveraging large-scale, pre-trained robotic datasets (such as RT-X) to reduce data dependency and integrating Large Language Models (LLMs) to enhance the robot's logical reasoning for complex, multi-step instructions.
Summary updated