Guided Streaming Stochastic Interpolant Policy, Puming Jiang★, Meiyi Wang, Kelvin Lin★, Ce Hao★, Harold Soh★, Robotics: Science and Systems (RSS)
Links:
Inference-time guidance is essential for steering generative robot policies toward dynamic objectives without retraining. However, existing methods are largely confined to chunk-based architectures, which must generate an entire action sequence before execution begins. This introduces high latency and prevents the policy from reacting to moving obstacles or test-time preferences mid-execution.
In this work, we analyze the time evolution of the value function via the Backward Kolmogorov Equation to rigorously derive the optimal guidance term for Stochastic Interpolants (SI). We instantiate this in the Streaming Stochastic Interpolant Policy (SSIP), which aligns generative evolution with the robot’s physical execution time to enable fast, reactive control. To support diverse deployment needs, we propose two complementary mechanisms: training-free Stochastic Trajectory Ensemble Guidance (STEG), which computes gradients on-the-fly for zero-shot adaptation, and training-based Conditional Critic Guidance (CCG), which amortizes inference for low latency. We validate the framework on Push-T, Robomimic simulations, and a real Franka Panda robot.
Resources
You can find our paper here.
Citation
Please consider citing our paper if you build upon our results and ideas.
Puming Jiang★, Meiyi Wang, Kelvin Lin★, Ce Hao★, Harold Soh★, “Guided Streaming Stochastic Interpolant Policy”, Robotics: Science and Systems (RSS)
@inproceedings{jiang2026guided, title={Guided Streaming Stochastic Interpolant Policy}, author={Jiang, Puming and Wang, Meiyi and Lin, Kelvin and Hao, Ce and Soh, Harold}, booktitle={Robotics: Science and Systems (RSS)}, year={2026} }
Contact
If you have questions or comments, please contact Puming or Harold.
