双臂人形机器人货架拣选系统 · WBCD 2026 季军 · 最佳创意末端执行器奖Bimanual Humanoid Shelf-Picking · WBCD 2026 2nd Place · Best Creative End-Effector Award

Competition Background

WBCD 2026 (What Bimanuals Can Do) is a robotics competition focused on bimanual humanoid manipulation. Track 1 required a Unitree G1 humanoid to transfer a heterogeneous set of objects from a three-level shelf into a transport cart within two 10-minute runs.

Each shelf level demanded a different whole-body posture:

• Top shelf: upright pick (+8 pts)
• Middle shelf: bent pick (+10 pts)
• Bottom shelf: crouched pick (+10 pts)

Scoring was unforgiving: only a completed pick-carry-place cycle counted, partial credit was zero, dropping an item during transfer cost −3, and each run’s score was multiplied by the declared autonomy level (×2 for remote teleoperation, ×4 for full autonomy).

Team NU Roboticists (Northwestern University) achieved:

• 🥈 2nd place overall
• 🏆 Best Creative End-Effector Design Award
• 14 successful transfers across two runs, zero gripper faults

System Overview

The system operates at two layers: the end-effector layer (how to grip) and the system layer (how to operate).

End-Effector Design

The Challenge

The object set spanned soft, crushable items (chip bags, cling wrap, plush toys) and rigid, slippery, heavy items (a full soda can, a tennis ball, a speed cube), all staged inside small bins. No single gripper geometry served the full range.

Four requirements conflicted directly:

• R1 Reach: The bottom shelf sits just above the floor; stock Dex1-1 fingers cannot reach it
• R2 Grip force on hard objects: Enough normal force to lift the soda can without it sliding out
• R3 Gentle grip on soft objects: No crushing, no slipping on smooth surfaces
• R4 Compactness: Fingers must fit inside small bins without knocking neighbors

Design Evolution

Baseline (stock Dex1-1) → Failed: Too short for the bottom shelf (fails R1).

Attempt 1: Widened finger → Failed: Width solved neither reach nor agility in tight bins.

Attempt 2: 3× length PLA finger → Failed: Reached the bottom shelf, but snapped at the base under the soda can load. Multiple revised versions all fractured.

The pivot — sleeve a glove, don’t replace the finger. Instead of redesigning the finger, keep the stock Dex1-1 fingers and design a TPU glove that slides over them like a fingertip cap:

• Long enough to reach the bottom shelf (R1)
• Compact cross-section for tight bins (R4)
• Curved tip for form closure on round hard objects (R2)
• TPU elastomer flexes repeatedly without fracturing (durability)

Mounting: From Exact Negative to Cylindrical Locators

The obvious approach — seeding the glove cavity from a Boolean negative of the finger — is intolerant of print variation: any slight over-extrusion prevents the glove from fitting at all. Instead, the cavity was relieved and cylindrical locating features (pins + through-rod) were added, constraining the glove at a few defined contacts rather than across an entire matched surface. This made mounting insensitive to print tolerance and reduced glove swaps to seconds.

Final Asymmetric Configuration

The counter-intuitive logic: the stiff hand is better for soft objects; the compliant hand is better for hard ones. Soft objects fail by crushing or squeezing out — a controllable stiff jaw avoids both. Hard round objects fail by sliding from a flat rigid pinch — a compliant curved tip wraps them into a stable seated contact. Each object goes to the hand whose failure mode it avoids.

Whole-Body Control: SONIC vs HOMIE

The team evaluated two whole-body control frameworks:

HOMIE is a model-predictive control approach with theoretically higher precision. In practice, it showed instability under the dynamic shelf-picking motions and required significantly more tuning time.

NVIDIA SONIC is a large motion-tracking policy that retargets the operator’s full-body pose directly to the G1. The operator wears a PICO4U headset and leg trackers and physically walks, bends, and crouches — the robot mirrors the motion. Balance is entirely the policy’s responsibility; the operator thinks only about where to be and what to grab.

SONIC was selected for:

• Intuitive mapping with minimal operator learning curve
• Stable performance across all three shelf postures
• Clean integration with the gripper server and camera pipeline

Active 2-DOF Head (TWIST2-Inspired Neck Design)

A fixed torso camera or wrist camera cannot simultaneously give a close-up grasp view and a view of the robot’s feet — yet the feet matter critically: a leg brushing the shelf rack during close approaches can tip the entire unit.

Inspired by the TWIST2 system, the team built a custom 2-DOF active head carrying a ZED Mini stereo camera:

• The head tracks the operator’s natural head motion (pitch and yaw) via our own ROS 2 control implementation
• The operator looks down — the robot looks down. No second camera needed.
• ZED video was tuned (resolution + DDS settings) to ~100 ms average latency to the operator laptop, relayed to the headset over TCP

Negative result — wrist camera: An Intel RealSense D405 wrist camera was tested for near-field depth during grasping. The G1’s onboard CPU could not absorb the additional load: latency rose from ~100 ms to ~400 ms, making teleoperation unusable. It was removed. The lesson: every sensor has a CPU cost; a sensor that degrades the primary video path has negative value regardless of the information it adds.

Torque-Aware Hold Mode

The operator commands the gripper via a controller trigger mapped to finger position, with no force or touch feedback. The raw interface has a critical failure mode: when fingers close on a stiff object, the commanded position keeps advancing while the actual position stalls, error accumulates, and the motor faults on over-torque. Holding a trigger at exactly the right partial depression for an entire transport walk is unrealistic.

The solution was implemented in the gripper motor server, running at a 2 ms control period:

• Normal: motor tracks trigger-mapped position via impedance control
• Detection: |estimated torque| > hold threshold AND |joint velocity| < stall threshold
• On trigger: latch current position as q_hold, switch to soft holding gains — position error collapses to zero, torque settles at a modest holding level
• Release: operator opens trigger past release margin, or load torque stays low for 300 ms

Effect: The operator squeezes fully and keeps holding. The server latches on first firm contact, over-torque faults disappear, and the grasp survives the entire transport walk without the operator metering force. Zero gripper faults across all competition runs.

What Failed

Results

Both runs were declared as remote teleoperation (×2 multiplier). The improvement in Run 2 came from operator familiarity, not hardware changes.

Zero over-torque faults across all runs. Every object class picked successfully, including the full soda can that had destroyed the rigid PLA finger.

Acknowledgments

This project was completed by Jyothi Swaroop Kasina, Chenyu Zhu, Andnet DeBoer, and Md Saif Ahmad, all from Northwestern University. All authors contributed equally. We thank the WBCD 2026 organizers for an outstanding competition platform.