ARMFUL: Learning Whole-Arm Grasp Configurations for Single-Arm Pickup of Large Objects
TL;DR A simulated dataset of 1.7 million whole-arm grasps on 793 large objects, each stress-tested by shaking and pushing, trains a diffusion model that chooses how one Franka arm wraps around a stool, box or rack to lift it. The pipeline lifts 85.4% of training and 69.9% of held-out objects in simulation, and 70 of 100 real trials on 10 IKEA objects.
What they do
Chairs, storage boxes and racks are too big for a parallel-jaw gripper, but a single arm can pick them up by wrapping its links around them. Such wraps are hard to teleoperate, and existing datasets cover almost none. The authors build WAG, a simulated dataset of whole-arm grasp configurations, and train ARMFUL on it to pick a wrap from a point cloud.
How it works
- Dataset. 7-DoF Franka joint configurations are sampled near each object in Isaac Lab, with penetrating ones discarded. Fixed squeeze (elbow and wrist) and lift (shoulder and upper-arm) joint offsets then test each one; the PD targets stall on contact, which produces the squeezing force.
- Stress test. A survivor and four jittered copies (±0.05 rad) must each hold the object through 5 cycles of ±0.3 rad base shaking at 1.1 Hz and a 30 N push. The result is 1.7M configurations on 793 objects, generated in about 3.5 days on one RTX 5090.
- Model. A DDPM over the 7 joint angles, conditioned on a DP3 point-cloud embedding and on which side of the base the elbow starts.
- Deployment. Samples are filtered for penetration and self-collision, then reached with RRT-Connect, with up to 10 draws. Real point clouds are fused from five depth cameras.
Results
- Simulation lift success is 85.4% on training objects and 69.9% on held-out ones. A mid-height wrap heuristic scores 54.6% and 54.3%; nearest-neighbor retrieval scores 58.0% on held-out objects.
- 96% (training) and 91% (held-out) of ARMFUL’s lifts survive the shake-and-push test, against 85% and 84% for the heuristic.
- Training without the stress test, on the same number of configurations, lowers lift success by 3.9 and 2.5 points.
- Real robot: 70/100 lifts on 10 unseen IKEA objects, 3/10 to 10/10 per object. In simulation at the same placements, the full pipeline given the exact object mesh gets 79/100; the heuristic gets 67.8% and retrieval 45.0%.
Limitations
Partial point clouds miss thin parts. Heavy, smooth objects slip, and fabric or bolted parts deform or rotate, which the rigid simulation with uniform mass and friction does not model. A single hand-designed squeeze-and-lift motion, with no regrasp, is baked into the dataset.
Videos
Figures



Abstract
Many everyday objects, such as storage boxes, baskets, and chairs, are too large or heavy for a gripper to grasp, yet a single arm can still pick them up by wrapping around and supporting them. Such grasps are hard to learn: wrapping motions are difficult to teleoperate, and existing datasets contain little whole-arm pickup data. We introduce WAG (Whole-Arm Grasp Dataset), the first large-scale dataset for single-arm pickup of large objects, with 1.7 million joint configurations across 793 objects. We sample configurations directly in joint space, execute each through squeeze and lift primitives in simulation, and stress-test survivors under active shaking and external forces. To demonstrate the utility of WAG and accomplish the whole-arm pickup task, we develop ARMFUL, a whole-arm pickup pipeline that pairs a diffusion model trained on WAG, generating whole-arm grasps from the object's point cloud, with a motion planner that reaches them collision-free before the same squeeze-and-lift execution. In simulation, ARMFUL reaches 85.4% lift success on training and 69.9% on held-out objects, and it lifts the object in 70 of 100 trials on a physical Franka arm across 10 IKEA objects.
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.
