Research Analyzer
← Back ICRA 2026

Viper: Verifiable Imitation Learning Policy for Efficient Robotic Manipulation

Xianfeng Cheng, Qing Gao, Guangyu Chen, Rui Xiong, Junjie Hu, Yulan Guo, Zhaojie Ju

PDF

AI summary

Key figure (auto-extracted from paper)
Viper unifies imitation learning with nonlinear model predictive control to achieve both high-precision tracking and real-time responsiveness in robotic manipulation.
Imitation Learning Model Predictive Control Robotic Manipulation Policy Stability Real-time Control Neural Dynamics

Problem

Existing robotic imitation learning methods struggle to balance motion precision with computational speed, as generative models either accumulate errors over long horizons or incur prohibitive inference latency.

Approach

Viper replaces online optimization with a learned multi-layer predictive model that forecasts optimal state-action sequences, integrated into a receding-horizon feedback loop with a Lyapunov-based stability guarantee.

Key results

  • NMPC-inspired framework that learns nonlinear dynamics and predicts optimal state-action sequences offline
  • Theoretical proof of practical stability for the closed-loop IL policy using a custom Lyapunov function
  • Reduced computational overhead via sparse visual penalties and static-dynamic encoding mechanisms
  • Validated high precision and fast execution across simulated and real-world manipulation tasks

Why it matters

Offers a theoretically verifiable and computationally efficient alternative to existing IL methods, advancing the deployment of precise, real-time robotic manipulation systems.

Abstract

Imitation learning (IL) presents a promising paradigm for enabling embodied robots to efficiently acquire human-like manipulation skills. However, prevailing methods face a persistent trade-off between motion precision and com- putational tractability. To resolve this fundamental challenge, this paper introduces Viper, a framework for Verifiable Imita- tion learning Policy for Efficient Robotic manipulation. Viper integrates principles of Nonlinear Model Predictive Control (NMPC) within a learning-based model. Grounded in an NMPC-style closed-loop architecture, the proposed method unifies the modeling of nonlinear system dynamics with online, multi-horizon optimization of state-action predictions, while intrinsically embedding physical constraints. This co-design enables both smooth trajectory generation and fast execution. Furthermore, a theoretical stability analysis for the Viper framework is provided. Extensive evaluations, from simulated benchmarks to real-world manipulation tasks, demonstrate that Viper effectively reconciles the competing demands of precision and speed inherent in existing robotic IL paradigms. Project page: https://cheng122.github.io/Viper

Index terms

Imitation Learning Learning from Demonstration

Related papers