Search papers, labs, and topics across Lattice.
This paper introduces a novel framework for grounding open-vocabulary responses in elephant-inspired soft-trunk robots, addressing the challenges of translating high-dimensional whole-body commands into feasible motions. By employing lightweight flow matching, the authors convert multimodal language model outputs into intent-intensity tuples and use Catmull-Rom spline controls for actuation trajectories. The framework significantly enhances grounding correctness from 25.0% to 77.2% and improves user satisfaction in human-robot interactions, demonstrating both efficiency and effectiveness in motion generation.
Grounding open-vocabulary commands in soft-trunk robots can boost user satisfaction ratings from 46% to 82% by enabling more expressive and feasible motions.
For close-contact human-robot interaction (HRI), trunk-like continuum manipulators provide a physical channel for diverse whole-body expression, but grounding open-vocabulary responses into such robots is difficult: end-effector motion underspecifies body shape, whereas direct whole-body commands are high-dimensional and hard to keep feasible. We propose a whole-body semantic-to-actuation grounding framework for elephant-inspired soft-trunk HRI based on lightweight flow matching. The framework converts responses from a multimodal large language model into bounded, morphology-aligned intent-intensity tuples, parameterizes tendon-actuation trajectories with compact Catmull-Rom spline controls, and uses a rectified-flow generator to sample feasible whole-body trunk motions. Experiments show that the proposed framework improves held-out grounding correctness from 25.0% to 77.2% over a raw-response dense-regression baseline. Compared with a denoising-diffusion baseline, it improves correctness from 71.9% to 77.2% and reduces inference time from 7.86 ms to 4.87 ms while preserving motion diversity. A 100-participant physical HRI study further shows that adding the generated soft-trunk motion channel increases the positive overall-satisfaction rating from 46% to 82% over the audiovisual-only baseline.
