Grasp-Taxonomy Conditioning as a Mode Switch for Dexterous Grasping of Climbing Holds

Rumi Loghmani, Abraham George, Amir Barati Farimani

November 2026

TL;DR BetaGrasp adds a four-class grasp-type label (crimp, sloper, pinch, jug) to a DP3-style point-cloud diffusion policy on a Franka arm with a LEAP hand. On 80 paired hardware trials on climbing holds, average slip force rises from 9.6 N to 14.3 N and full slip-offs drop from 44% to 10%.

What they do

A dexterous hand has to grasp different objects with different hand shapes, but a diffusion policy trained on several grasp types gets nothing in its input that says which one to use. BetaGrasp gives a DP3-style point-cloud policy a four-class grasp label (crimp, sloper, pinch, jug) and tests it on 3D-printed climbing holds with a Franka arm and a 16-DoF LEAP hand. A spring testbed with a ratchet turns each trial into a slip force in newtons.

How it works

Policy. A PointNet-style encoder (1024 XYZ points from four RealSense cameras), a proprioception MLP and a small grasp-label MLP are fused into a 512-d vector that conditions a 1D temporal U-Net denoiser. It outputs 23-dim joint-space action chunks at 10 Hz with 10 DDIM steps. The baseline is the same network without the label branch.
Data. 200 VR-teleoperated demonstrations, 50 per grasp type across five hold orientations (46.4 minutes). The operator supplies the label in training and at deployment.
Testbed. The hold sits on a deck that slides along one axis against two springs. After the grasp the arm pulls 13 cm, and an 11-tooth ratchet records the peak displacement, giving slip force in steps of about 2.6 N.

Results

Simulated closed-loop rollouts from the reset pose: BetaGrasp reaches the correct grasp type in 94% of rollouts, the baseline in 25%, because all 800 of its rollouts end in a jug grasp.
The point-cloud branch alone does not separate the four holds (5-NN accuracy 39%, chance 25%).
Hardware, 80 paired trials: average slip force 14.3 N against 9.6 N. BetaGrasp wins 61 pairs, ties 14 and loses 5. The largest gain is on pinch (+6.8 N), the smallest on sloper (+3.0 N), where the baseline slipped off in all 20 trials.
Grasps that slip off before moving the hold drop from 44% to 10%.

Limitations

One embodiment, one operator and one physical hold per grasp type, with no held-out holds, so part of each gain may come from the particular hold. The label is supplied by a person. The testbed loads the grasp along one fixed axis, the rollout analysis is simulated, and 14.3 N is a moderate grip.

Figures

From Figure 1: BetaGrasp architecture. Point cloud, proprioception and grasp-type label are encoded independently and... — From Figure 1: BetaGrasp architecture. Point cloud, proprioception and grasp-type label are encoded independently and fused into a 512-d conditioning vector for a 1D temporal U-Net diffusion denoiser. The DP3 baseline drops only the grasp-type branch (orange), 0.1% of the 34.6 M parameters.

From Figure 2: Spring-testbed benchmark. The LEAP Hand v1 (16-DoF, on a Franka Emika Panda) executes a taxonomy-condi... — From Figure 2: Spring-testbed benchmark. The LEAP Hand v1 (16-DoF, on a Franka Emika Panda) executes a taxonomy-conditioned grasp on a 3D-printed jug hold seated on the testbed.

From Figure 3: Final hand poses of simulated closed-loop rollouts started from the deployment reset pose, projected o... — From Figure 3: Final hand poses of simulated closed-loop rollouts started from the deployment reset pose, projected onto the first two principal components of the demonstrators’ final hand poses (light gray; marker shape = hold). BetaGrasp rollouts (orange) mostly land on the demonstrated grasp for their hold; DP3 rollouts (gray outline) end in the jug grasp on every hold.

From Figure 4: Slip force per hold on the spring testbed, 20 trials per policy per hold. Boxes show the interquartile... — From Figure 4: Slip force per hold on the spring testbed, 20 trials per policy per hold. Boxes show the interquartile range and median; dots are individual trials. The 5.2 N floor is the spring preload, reached when the grasp slips off before the hold moves.

Abstract

Dexterous grasping in robotics relies on distinct hand configurations depending on object geometry. We introduce BetaGrasp, a 3D diffusion policy that conditions a DP3-style point-cloud policy on a four-class grasp label (crimp, sloper, pinch, and jug), running on a Franka arm with a 16-DoF LEAP hand. We also introduce a climbing-hold benchmark: a spring testbed whose ratchet gives a slip-force reading in newtons. In our analysis the policy's point-cloud encoder did not separate the four holds, so the scan alone gave it little to choose a grasp from. Without the label (a DP3-style baseline), simulated rollouts from the robot's reset pose all end in a jug grasp, whatever the hold (25% correct grasp type, versus 94% for BetaGrasp). The label works as a switch for which grasp the policy performs. On 80 paired hardware trials, BetaGrasp raises average slip force from 9.6 N to 14.3 N, and grasps slipping fully off the hold drop from 44% of trials to 10%.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong