Rostok Framework: Automating Underactuated Robot Gripper Design

TL;DR – A new open‑source tool, Rostok, generates tendon‑driven grippers by treating the robot’s body as the primary source of behavior. The pipeline runs a series of optimization and graph‑grammar steps, produces designs that score highly on a simulated reward function, and then validates the best candidates on a physical test rig.

---

The problem: designing compliant grippers is hard

Underactuated grippers rely on flexible linkages and tendons to adapt to objects of varying shape. Traditional design cycles involve hand‑crafted linkage geometry, extensive trial‑and‑error, and costly prototyping. Engineers must balance competing criteria: range of motion, force transmission, manufacturability, and the ability to hold objects securely without active control. The process is especially painful for small research labs that lack dedicated mechanical design teams.

The Rostok approach

Rostok tackles the problem from two angles:

1. Morphological computation – Instead of programming a controller to achieve a grasp, the framework shapes the robot’s morphology so that the desired motion emerges naturally from the passive dynamics of the mechanism.
2. Generative design pipeline – A combination of graph‑grammar synthesis, heuristic search, and physics‑based simulation explores thousands of linkage topologies. Each candidate receives a reward based on simulated grasp success, energy efficiency, and manufacturability.

The pipeline is broken into clear stages:

Stage	What happens
Grammar expansion	A graph‑grammar ruleset creates a library of possible linkage graphs.
Parameter sampling	For each graph, joint lengths, tendon routing, and spring constants are sampled within realistic bounds.
Simulation	The sampled mechanisms are evaluated in a physics engine (Bullet) that models tendon forces and contact friction.
Reward scoring	A weighted function combines grasp stability, required actuation torque, and a penalty for parts that are difficult to 3‑D print.
Selection & refinement	Top‑scoring designs undergo local gradient‑based refinement to fine‑tune dimensions.
Real‑world test	The final set is fabricated (mostly via SLA printing) and mounted on a test rig that measures actual grasp forces and repeatability.

All of these steps are orchestrated by Python scripts that call into the open‑source libraries rostok, pybullet, and trimesh. The code lives on GitHub at https://github.com/aimclub/rostok and includes example configuration files for reproducing the results presented in the paper.

How the generated grippers perform

The authors report that the top three designs achieved a simulated grasp success rate of 92 % across a benchmark set of 20 objects ranging from cylinders to irregularly shaped tools. When transferred to the hardware setup, the physical grippers retained 85 % of that success rate, a drop that the authors attribute to minor tolerances in 3‑D printing and unmodeled friction.

Key performance numbers from the paper:

• Maximum holding force: 12 N (average across objects) – comparable to commercially available soft grippers that cost an order of magnitude more.
• Actuation torque: 0.35 Nm per tendon – low enough for a single hobby‑grade servo to drive the whole hand.
• Fabrication time: ~2 hours per prototype, including printing, cleaning, and tendon routing.

These results suggest that the framework can produce viable gripper designs with a fraction of the engineering effort normally required.

Why this matters for the robotics community

1. Lower barrier to entry – Researchers without a dedicated mechanical shop can now generate functional grippers from a laptop and a desktop 3‑D printer.
2. Rapid iteration – By automating the exploration of the design space, Rostok shortens the time from concept to physical test from weeks to days.
3. Open ecosystem – Because the source code, data sets, and simulation environments are all released under a permissive license, other groups can extend the grammar, add new actuation models, or integrate the pipeline with reinforcement‑learning controllers.
4. Potential for specialization – The reward function is modular, allowing developers to prioritize traits such as ultra‑lightweight structures for aerial robots or high‑force grips for industrial pick‑and‑place.

Limitations and next steps

The current implementation assumes planar linkage mechanisms and does not yet support fully 3‑D articulated structures. Also, the simulation relies on simplified tendon models that ignore hysteresis, which can become significant for high‑speed applications. Future work outlined by the authors includes:

• Extending the grammar to generate spatial linkages and multi‑finger hands.
• Incorporating more sophisticated tendon dynamics (e.g., cable stretch, friction).
• Adding a closed‑loop control layer that can exploit the passive compliance while providing fine‑grained force modulation.

Community reaction

Since the pre‑print appeared on arXiv, the repository has attracted over 800 stars and several forks that add custom objective terms for tactile sensing and modularity. A handful of startups have already begun experimenting with Rostok‑generated grippers in prototype assembly lines, citing the speed of design as a decisive factor.

---

For a deeper dive into the technical details, see the full paper (CC BY 4.0) linked in the article and the accompanying GitHub repository.