Dice Question Streamline Icon: https://streamlinehq.com

Controllable mode selection in TraceGen’s stochastic-interpolant decoder

Investigate and develop alternative stochastic-interpolant schedules and explicit control mechanisms within the TraceGen flow-based trace decoder—beyond the current linear interpolation ODE—to enable selection of a desired trajectory mode when multiple plausible futures exist for ambiguous manipulation tasks.

Information Square Streamline Icon: https://streamlinehq.com

Background

TraceGen generates future scene-level 3D trajectories using a flow-based decoder trained under the Stochastic Interpolant framework, instantiated with a linear interpolation ODE schedule. This choice enables sampling diverse outputs by varying noise seeds but does not provide explicit control over which trajectory mode is produced in multimodal settings.

The authors explicitly note they have not explored different interpolant schedules or mechanisms for mode control. Addressing this gap would allow deliberate selection among multiple plausible futures—crucial for ambiguous tasks where several valid motion outcomes exist.

References

Within the Stochastic Interpolant framework, we adopt linear interpolation with ODE integration. While this approach allows sampling diverse trajectories through different noise initializations, we have not yet explored alternative interpolation schedules or mechanisms to explicitly control which trajectory mode is generated for ambiguous tasks.

TraceGen: World Modeling in 3D Trace Space Enables Learning from Cross-Embodiment Videos (2511.21690 - Lee et al., 26 Nov 2025) in Section 7: Limitations and Future Work