1Biomimetic Robotics Lab, Massachusetts Institute of Technology
23D Vision Lab, Seoul National University
3RoPhi Lab, Yonsei University
In this work, we study whether a robotic hand using proprioception alone can grasp diverse objects with no visual observation. We present a modular dexterous grasping architecture that separates global arm motion from local contact control. An independently controlled arm guides the hand toward the object, while a reinforcement learning policy grasps and stabilizes it using only hand proprioceptive feedback. We call this a blind grasp reflex: grasping without images, object poses, or geometric observations.
A learned stable-grasp score determines when the object is securely held, allowing the arm to begin post-grasp manipulation. This separation makes grasping a reusable hand-level skill that can be combined with independently designed arm controllers for various manipulation tasks. Experiments in simulation and on hardware demonstrate robust blind grasping across diverse objects and seamless composition with a range of arm controllers. Moreover, despite never observing contact geometry, the learned grasp score closely aligns with an independent physics-based measure of grasp stability. The resulting approach follows a simple principle: see to reach, feel to grasp.
Our framework separates global arm motion from local grasp control. The arm module sends desired palm poses and a grasp-or-release command. The learned hand module returns finger targets and a scalar grasp score that tells the arm when to begin manipulation.
Explore recorded simulation rollouts from different viewpoints. The arm module guides the palm to the initial object position, and when it is within range it gives the grasp enable command to the hand. Then the hand policy begins grasping and when it reports a high enough grasp score the arm module lifts the palm.
Twelve geometrically diverse grasps were selected from the full YCB evaluation set, spanning fruit, a ball, containers, tools, and household objects. Each recording is disturbance-free and shows more than five seconds of continuous object hold. The rollouts render directly in your browser; no live simulation server is required.
The blind grasp reflex generalizes across object sets and responds robustly when the object moves after execution begins.
| Evaluation | Ours | End-to-end RL |
|---|---|---|
| YCB objects grasped (78 objects) | 96% | 98% |
| GraspXL objects grasped (3,028 objects) | 95% | 97% |
| Dynamic grasp success | 92% | 8% |
| Object-pose generalization | 90% | 15% |
The learned grasp score achieves a mean within-episode Spearman correlation of +0.77 with the Ferrari-Canny force-closure metric, even though the deployed policy never observes the contacts used to compute it.
The following videos show the learned grasp signal, transfer across robot configurations, blind grasps behind occluders, reactive grasping, lightbulb screwing, and language-conditioned manipulation. Every demonstration uses the same frozen hand policy.
A simulation rollout shown alongside the learned grasp score and the independent Ferrari-Canny force-closure metric. Both signals rise as the grasp forms, fall after an external disturbance, and recover during re-grasping.
The policy observes only the hand, so the same checkpoint can be reused across arm kinematics and mounting configurations without retraining.
The same hand policy grasps with a Flexiv Rizon 4, Franka Panda, UR5e, and Kinova Gen3 in simulation.
The policy remains effective with the arm mounted on the floor, side, wall, and ceiling - configurations not encountered during training.
We show that our method truly relies on feel rather than sight by asking a VLM to grasp an object hidden behind an occluder.
The hand searches behind the box and secures an object hidden from the camera.
The robot reaches past the visible cup and grasps the occluded target behind it.
An open laptop fully blocks the target while the hand acquires it through contact.
The hand grasps a pair of soccer shoes while the ball obscures the contact region.
The robot retrieves a measuring tape hidden behind the helmet.
The hand finds and lifts a hidden toy in a cluttered tabletop scene.
Here, another arm controller uses SAM 2 to track objects in real time while the hand runs the same frozen blind-grasping policy. State-of-the-art reinforcement-learning grasping methods typically learn arm and hand control together, limiting the entire system to behaviors seen during training. Our method separates the two, enabling reliable grasping while the arm is controlled by any module—not only the reach-and-lift controller used during training.
A VLM-guided pick-and-place controller directs the arm while the hand runs the blind, proprioception-only grasping policy. The arm approaches the pickup position supplied by the VLM and, once within range, instructs the hand to begin grasping. Only after the hand returns a sufficiently high grasp score does the arm proceed to the placement position, while the hand maintains the grasp throughout the motion. The arm then instructs the hand to release, placing the object. These demonstrations show the hand policy grasping objects at heights never encountered during training and producing both power and pinch grasps across a diverse set of objects.
Sequential commands place a sloth toy, a cheetah toy, and a gelatin box into the blue bin.
The system places tape in a bin, a yellow switch in a helmet, and an orange voltage meter on a white box.
Natural-language instructions place a pumpkin on a plate, a yellow apple in the green bowl, and a green apple in the blue bowl.
The arm begins rotating only after the hand's learned grasp score indicates that the lightbulb is secure. The hand remains active during sustained wrist motion, then releases between turns so the screwing sequence can repeat.
We qualitatively evaluate the system on 29 real-world objects spanning toys, tools, containers, deformable items, and irregular shapes. The same policy successfully grasps every object shown below.
Blind grasps across compact toys, tools, and angular objects with varied contact geometry.
The same hand policy stabilizes spherical and cylindrical objects spanning several sizes and surface properties.
Hardware trials with bottles, cartons, and packaged goods demonstrate reliable contact-driven grasp acquisition.
The hand adapts online to deformable and asymmetric objects without observing their shape or contact geometry.
