cs.ROApr 20, 2025
SaveSoft yet Effective Robots via Holistic Co-Design
Organizations: Laboratory for Information & Decision Systems, Massachusetts Institute of Technology, Cambridge, 02139, MA, USA. · Cognitive Robotics, Delft University of Technology, Mekelweg 2, Delft, 2628 CD, Netherlands. · Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, 02139, MA, USA. · The BioRobotics Institute, Scuola Superiore Sant’Anna, Pisa, 56025, Italy. · Embodied AI AG, Lausanne, 1015, Switzerland. · CREATE Lab, EPFL, Lausanne, 1015, Switzerland. · Advanced Robotics Centre, Department of Mechanical Engineering, National University of Singapore, 117575, Singapore.
Abstract
Soft robots promise inherent safety via their material compliance for seamless interactions with humans or delicate environments. Despite progress, the field struggles to balance task-specific performance with broader factors like durability and manufacturability--a difficulty that we find is compounded by traditional sequential design processes with their lack of feedback loops. In this perspective, we review emerging co-design approaches that simultaneously optimize the soft robot's body and brain, enabling the discovery of unconventional designs highly tailored to the given tasks. Their adoption is limited by narrow objectives, gaps between simulated and real performance, and computational cost. To address these challenges, we propose a holistic co-design framework that incorporates a broader range of design values, integrates real-world prototyping to refine evaluations, and boosts efficiency through surrogate metrics and model-based control strategies. Finally, we outline research priorities in design priors and metrics, AI-assisted evaluation, and balancing computational refinement with physical testing and safety with performance.
Figures & tables
Figure 1: Holistic Co-Design of Soft Robots. The five pillars of holistic co‑design are (1) Incorporating comprehensive design requirements and values directly into multi‑objective optimization; (2) Efficiently exploring the full design space by melding design priors (e.g., biological inspirations, existing solutions) with computationally efficient co‑design routines; (3) Explicitly accounting for design realization (e.g., prototyping and testing) via a probabilistic treatment of evaluation metrics and by formalizing the refinement‑vs‑realization trade‑off—enabling targeted prototyping to reduce metric uncertainty and narrow the sim‑to‑real gap; (4) Ensuring auditability to preserve design knowledge and guarantee reproducibility. (5) Fostering cross‑disciplinary collaboration and involving all relevant stakeholders in defining design values and providing iterative feedback; These pillars are inherently interconnected (see green arrows)—for example, stakeholders and the design team co‑define design requirements, certain design values (e.g., HRI) demand evaluation in the real world (i.e., enabled by realization), and both the exploration of the design space and the fulfillment of design requirements remain fully auditable.
Figure 2: The traditional soft robot design process vs. holistic co-design. Left: A traditional sequential design cycle van Boeijen et al. (2020) applied to soft robots. This sequential workflow lacks iterative feedback loops, leading to information silos and preventing regular, bidirectional sharing of insights and data across the development team and the relevant stakeholders. Right: A holistic co-design framework that complements computational refinement with prototyping across various levels of design readiness - visualized as “onion” layers in this graphic. Solid yellow lines denote the prototyping pathway, while dashed orange arrows illustrate how empirical findings continuously inform and update the estimates for the real‑world design values .
Figure 3: A Framework for Efficient Computational Co-Design of Soft Robots. We sample reduced-order design parameters from an initial distribution that can include design priors from biological inspiration Mazzolai and Laschi (2020) ; Chen and Wang (2020) ; Laschi and Mazzolai (2024) , known mechanisms, or existing soft robot designs. This design space—either learned or explicitly defined by physical or geometric values (i.e., size optimization)—is translated by a design decoder into a detailed robot body description (such as a 3D mesh with sensor and actuator placements) that provides immediate manufacturability feedback. A reduced-order model is then derived Alora et al. (2023) ; Stölzle and Della Santina (2024) ; Valadas et al. (2025) ; Alkayas et al. (2025a) or learned to efficiently assess workspace, open-loop compliance, and embodied intelligence Cianchetti (2021) ; Mengaldo et al. (2022) ; Vihmar et al. (2023) , offering inexpensive feedback to the optimizer. Similarly, perception and control systems can be developed using this model (via model-based control Della Santina et al. (2023) , MPC Alora et al. (2023) , or differentiable physics Spielberg et al. (2019) ; Wang et al. (2023b) ; Stölzle et al. (2026a) ), while observability and controllability are evaluated without costly simulations. Finally, closed-loop simulations (low-fidelity using the reduced-order model or high-fidelity FEM-based Coevoet et al. (2017) ) assess integrated performance metrics, which in turn optimize the sampling distribution, decoder, and all relevant system parameters.
Figure 4: Refinement vs. Realization. Top Left: This panel shows how refinement affects the design sampling distribution . Unlike (most) conventional co‑design methods that sparsely explore the design space, we condition on design priors (e.g., biological inspiration and existing soft robot designs) and iteratively update it via the co‑design optimizer until the optimum is reached. Bottom Left: Here, we demonstrate how iterative realization refines probabilistic evaluation metrics. Rather than using deterministic metrics that ignore inherent simulation uncertainty, we define a probabilistic distribution for each metric, with purposeful prototyping reducing the uncertainty over iterations. Right: This panel examines the tradeoff between refinement and realization. We define the realization ratio as the fraction of realization iterations relative to the total iterations and the sim‑to‑real error as . We then plot the evolution of real-world cost versus sim‑to‑real error as iterations increase for a fixed : when , designs are optimized computationally with realization occurring only at the end, while means only the confidence in the evaluation metrics is increased. The optimal ratio efficiently minimizes the real-world cost to its optimum via refinement and reduces the sim‑to‑real error to zero by updating via realization.
Supplementary Fig. 1: Illustration of the proposed safety metric for quantifying the injury severity of a local blunt contact between a soft robot and a human body part. The maximum contact pressure/stress is used as a proxy for injury risk International Organization for Standardization (2016) . The contact force is obtained from a one-dimensional projected collision model along the contact normal , with penetration coordinate . To obtain a conservative estimate, the human-side motion is prescribed, corresponding to the limiting case in which the human body part is effectively constrained or has inertia much larger than the reflected robot inertia .
| Metric | Experimental protocol | Computational implementation |
|---|---|---|
| Pick success rate | Count successful picks over total attempts in greenhouse trials. | Simulate grasp dynamics and apply success probability models under domain-randomized variability. |
| Selectivity error (ripe vs. unripe) | Compare picked fruits with human-annotated ripeness ground truth. | Classifier accuracy on synthetic or augmented datasets representing different ripeness states. |
| Cycle time | Measure time per pick from actuation start to fruit release. | Model task scheduling with actuator and robot dynamics; stochastic timing models. |
| Throughput per hour | Number of fruits successfully harvested per hour in pilot trials. | Extrapolation from cycle time distributions under randomized task sequences. |
| Dropped fruit percentage | Ratio of fruits lost during transfer to total picked fruits. | Simulated grasp stability analysis with randomized external perturbations. |
| Bruise percentage (24/48h) | Visual inspection and pressure mapping after storage at 24/48 h. | Contact force simulation with tissue deformation models and probabilistic damage thresholds. |
Supplementary Table 1: Probabilistic evaluation metrics with experimental protocols and computational implementations.
| Sensor modality | Function | Example commercial sensors | Typical algorithms |
|---|---|---|---|
| Multi-view RGB-D / stereo | 3D fruit pose, occlusions, approach planning | Intel RealSense (e.g., D455); Stereolabs ZED2i; industrial RGB + structured light | Instance segmentation + tracking; depth fusion; 6D pose refinement |
| NIR / hyperspectral | Ripeness/defect cues under lighting variability | Specim FX10/FX17; multispectral cameras (agri/food) | Spectral features + classifier; domain adaptation; calibrated confidence |
| Close-up stem/peduncle vision | Detachment-point localization and cut/twist planning | Eye-in-hand RGB(-D); structured light | Keypoints/skeletonization; stem tracing; local surface normal estimation |
| Wrist force/torque | Contact monitoring, detachment detection, safety limits | ATI Mini45 (or similar); integrated F/T at flange | Contact-state estimation; impedance/admittance control; force thresholds |
| Tactile / fingertip sensing | Slip/contact patch; bruise-risk proxies | GelSight-type tactile; capacitive/optical tactile arrays | Slip detection; contact-area/pressure proxy; tactile servoing |
| Bin fill estimation | Packing density, drop/impact risk mitigation | Overhead RGB-D/stereo; depth camera | 3D occupancy/fill-level estimation; anomaly detection |
Supplementary Table 2: Perception stack for indoor tomato harvesting: modality, role, example sensors, and typical processing algorithms.