CRAFT: A Tendon-Driven Hand with Hybrid Hard-Soft Compliance

What Happened

Researchers have introduced the CRAFT hand, a novel anthropomorphic robotic hand that uses tendon-driven actuation combined with a hybrid hard-soft compliance system. The key insight is that contact forces during manipulation are not distributed evenly across the hand—impacts tend to concentrate at joints. By strategically placing hard structural elements at joints and softer, compliant materials elsewhere, the design achieves a balance between durability and adaptability. The tendon-driven mechanism mimics human muscle-tendon systems, enabling finer control while reducing mechanical complexity compared to fully motorized joints.

Why It Matters

This work addresses a fundamental tension in robotic manipulation: rigid hands offer precision but struggle with contact-rich tasks, while soft hands excel at adaptation but lack strength and repeatability. CRAFT’s hybrid approach is pragmatic—it does not sacrifice structural integrity where it is needed most (joints) while still allowing the rest of the hand to deform and conform to objects. For AI practitioners, this has direct implications for reinforcement learning and policy deployment. A hand that can absorb unexpected impacts without damage reduces the need for conservative control policies, potentially accelerating training in simulation-to-real transfer. Additionally, the tendon-driven architecture simplifies the kinematic model, making it easier to integrate with learning-based controllers that require accurate forward and inverse dynamics.

Implications for AI Practitioners

First, the CRAFT hand’s compliance characteristics could lower the sample complexity of manipulation policies. When a gripper can passively adapt to object geometry, the policy does not need to learn precise force trajectories for every interaction—the hardware handles part of the compliance. This is especially relevant for contact-rich tasks like assembly or in-hand manipulation, where traditional rigid grippers require meticulous force control.

Second, the tendon-driven design introduces a nonlinear actuation model that may require careful calibration when used with neural network policies. Practitioners should account for tendon stretch, friction, and hysteresis in their simulation environments, or risk sim-to-real gaps. The paper’s emphasis on “hybrid hard-soft compliance” suggests that the hand’s behavior is not purely passive—it can be actively modulated, which opens the door for adaptive compliance policies that trade off stiffness and dexterity on the fly.

Finally, this work reinforces a broader trend in embodied AI: hardware and policy design are increasingly co-optimized. Rather than treating the robot hand as a black box, researchers are building physical structures that encode prior knowledge about manipulation. For AI teams, this means that investing in hardware that “bakes in” compliance can reduce the burden on learning algorithms, especially in safety-critical or high-contact environments.

Key Takeaways

• CRAFT’s hybrid compliance design strategically places rigid materials at joints and soft materials elsewhere, balancing durability with adaptability for contact-rich tasks.
• The tendon-driven actuation simplifies control while enabling finer manipulation, but introduces nonlinearities that must be modeled accurately for sim-to-real transfer.
• For AI practitioners, this hardware reduces the need for conservative policies and may lower sample complexity in reinforcement learning for manipulation.
• The work exemplifies a broader shift toward co-designing robot hardware and learning algorithms, where physical compliance complements—rather than complicates—policy learning.