Answers are grounded in this paper and cite their sources.
Recent learning-based robot policies, including multimodal foundation models, have demonstrated remarkable progress in robotic manipulation and instruction-following by leveraging large-scale demonstration data and high-capacity architectures. These approaches promise general and flexible robot behavior across a wide range of tasks. However, their successful deployment on real robotic systems remains challenging. In particular, robot learning is fundamentally constrained by the scarcity and cost of collecting high-quality demonstration data, especially for physical manipulators. Moreover, large-scale models often require substantial computational resources, introducing latency and limiting their compatibility with real-time, high-frequency feedback control on embedded robotic platforms. As a result, lightweight and specialized models remain a highly practical and effective solution for many real-world robotic applications. Smaller architectures can be trained with limited data, enable fast and reliable inference, and can be deployed directly on robotic hardware while maintaining tight control-loop requirements. Importantly, such models do not sacrifice the advantages of modern sequence modeling approaches, but instead allow efficient specialization to specific control problems, where data efficiency, stability, and deployability are critical. In this paper, we present LiteHand, a lightweight Transformer-based policy trained from a small number of human demonstrations and enhanced through data augmentation to improve robustness and generalization. The proposed model learns to autoregressively predict incremental end-effector displacements conditioned on trajectory history, current robot state, and task context. We demonstrate the effectiveness of this approach on the non-trivial task of generating smooth, continuous handwriting trajectories for a robotic manipulator in closed-loop control. This task serves as a representative benchmark for precise, real-time trajectory generation. We evaluate LiteHand in both simulation and real robot experiments, showing that it achieves high trajectory accuracy, strong robustness to execution noise, and reliable generalization across varying workspace configurations. Crucially, the model enables stable closed-loop control at 50 Hz with low inference latency, demonstrating that compact Transformer policies trained with limited demonstrations can provide an effective, data-efficient, and deployable solution for real-time robotic control.
No takes yet. Share an insight, caveat, or question.
Maccarini et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: