In this work we study if 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. Project page: https://blindgraspreflex.github.io.
Figures & tables
Fig. 1: The proposed method can grasp a diverse set of real world objects blindly without visual observations and priors.
Fig. 2: Framework overview. A proprioceptive hand policy coordinates with the arm through an engage command and a learned grasp score, enabling reuse with different arm controllers without retraining.
Term
Definition
Purpose
Coverage
∑iwi1[ℓi]
Inner-hand contact
Height
clip(zt−zrest,0,zmax)
Object lifting
Hold
1[∥ptobj−pgoal∥<dtol]
Stable hold
TABLE I: Task reward terms.
Objects grasped (30 s)
Per-attempt
Dynamic
Object pose
Arm
Re-grasp
Method
YCB ( n=78 )
GraspXL ( n=3028 )
success
success
generalization
agnostic
ability
Scripted
31%
39%
39%
15%
30%
✓
✗
RobustDexGrasp [ 10 ] †
98%
97%
85%
8%
15%
✗
✓
Ours (no grasp score)
73%
82%
82%
75%
70%
✓
✗
Ours
96%
95%
85%
92%
90 %
✓
✓
TABLE II: Comparison against baselines. The end-to-end policy leads on static grasping, but collapses on dynamic objects and covers only part of the workspace; our method is within a few points of it on static success while retaining both.
Fig. 3: Grasp success as a function of where the object rests on the table ( x , y in metres from the robot base), for our method and the end-to-end baseline. The arm starts each episode at its home pose, and both policies saw the object only inside the training spawn box (orange).
Fig. 4: a) The VLM pick-and-place policy running on the arm moving tape from the top of the oscilloscope to the bin. b) The policy to screw in a lightbulb. c) Dynamic grasping where the object location and target location are both changed while the policy is running.
Fig. 5: 29 real-world objects that our method grasped during testing.
Fig. 6: Learned grasp score (left) and Ferrari–Canny ϵ (right) averaged across grasp attempts. The score is never supervised and never observes the contacts ϵ is computed from, yet the two separate at the same instant and decay together on failures; ρˉ is the within-episode Spearman correlation between them, averaged over episodes.
Fig. 7: (Top) Arm successfully grasping with different mounting positions not seen during training. (Bottom) Cross-embodiment across three different robot arms not seen during training.
Fig. 8: At deployment, QP is solved to compute joint velocities that satisfy the task space velocity command and also avoid collisions with obstacles in real-time.
Success
Time to hold (s)
ρˉ
Action space (PCA)
β=0
91%
3.8
0.72
β=0.5
96%
3.9
0.77
β=1
28%
3.9
0.70
Student observation
q
91%
4.1
0.74
TABLE III: Ablation results. Each block varies one design choice. bold indicates the design choice of the proposed method.
University of Michigan, Ann Arbor, MI, USA. · Department of Mechanical Engineering, MIT, Cambridge, MA, USA. · Department of AI, Yonsei University, Seoul, South Korea.