Scaling SoftAct to Longer-Horizon, Multi-Stage Manipulation

Develop an extension of the SoftAct contact- and force-aware retargeting and policy learning framework that can handle longer-horizon, multi-stage manipulation tasks in which contact modes and the functional roles of contacts evolve over time, while maintaining reliable performance with a non-anthropomorphic pneumatic soft robot hand.

Background

SoftAct transfers human manipulation skills to a non-anthropomorphic soft robot hand by explicitly leveraging contact geometry and force distributions instead of kinematic correspondence. The presented experiments demonstrate effectiveness across several contact-rich tasks but focus on relatively short-horizon behaviors with fixed interaction structure.

The authors note that many real-world manipulations are composed of multiple stages where contact modes and the functional roles of contacts change over time. Extending SoftAct to reliably operate in such longer-horizon, evolving-contact scenarios is identified as an unresolved challenge, motivating research into temporal abstraction, contact-mode reasoning, and robust policy execution over extended horizons.

References

Scaling the framework to longer-horizon, multi-stage manipulation remains an open challenge, particularly in settings where contact modes and functional roles evolve over time.

— Functional Force-Aware Retargeting from Virtual Human Demos to Soft Robot Policies  (2604.01224 - Yoo et al., 1 Apr 2026) in Section: Limitations and Future Work

These results support a compact structured contact correction in one frozen regime, not nonlinear speedup or transfer across Newton states; evolving contact, friction, and contact-rank and mesh scaling remain open tests.

— TOGEARI: Interaction-Space Preconditioning for Condensed Finite-Element Systems with IPC Contact  (2608.14162 - Liu et al., 14 Aug 2026) in Abstract; Sections 1 and 6, Conclusion

However, an MPC pipeline has not been successfully shown in prior work on deformable object tasks as complex as ours and with multiple behaviors to even search on top of and chain them, and we were unable to get it to converge on any plausible results. While the forward model worked, it was not useful as the MPC action sampling itself was the major bottleneck.

— DefFlow: Goal-Conditioned Deformable Object Manipulation using 3D Point Flows  (2610.22827 - Singh et al., 16 Oct 2026) in Footnote in Section 3, “Baselines”