WBCD: Dual-Arm Robotics Coordination

Updated 9 July 2026

WBCD is a framework that explores dual-arm coordination by benchmarking competitive and empirical tasks such as packing and table service.
It employs diverse methodologies including tactile manipulation, role-decomposed imitation learning, and vision-language adaptation for robust dual-arm control.
Evaluation metrics and benchmarks highlight progress and challenges in precision contact, long-horizon memory, and adaptive role exchange in dual-arm systems.

What Bimanuals Can Do (WBCD) denotes both a concrete competition setting at the 2025 IEEE International Conference on Robotics and Automation and a broader empirical program in dual-arm robotics concerned with the operational envelope of coordinated two-manipulator systems. In the competition literature, WBCD is instantiated through logistics packing and table service under explicit timing, autonomy, and transfer constraints; in the surrounding research literature, the same question is operationalized through tactile dual-arm manipulation, role-decomposed imitation learning, in-hand reconfiguration, vision-language one-shot adaptation, language-model in-context control, and long-horizon coordination benchmarks. Across these settings, the defining problem is not merely increasing the number of end effectors, but managing coupled state-action spaces, asymmetric and symmetric coordination, contact stability, hand-state continuity, disturbance recovery, and embodiment transfer (Li et al., 6 Jun 2025, Li et al., 20 Aug 2025, Lin et al., 2023, Peng et al., 7 Apr 2026).

1. Competition framing and official task structure

In the logistics setting documented by NeSyPack, the WBCD packing task is specified as

$T=[(O_1,N_{O_1}),\ldots,(O_N,N_{O_N}),S],$

where $O_i$ denotes an object category, $N_{O_i}$ the required count, and $S$ the final box-sealing operation. The physical setup comprises four storage bins containing known or novel objects, one central packing box, and a bimanual platform with two Galaxea A1X 7-DoF arms, each equipped with a one-finger ALOHA gripper and a wrist-mounted RealSense D435i depth camera. The task flow is sequential: identify which object category to pick next and from which bin, execute a bimanual pick with optional handoff or “stir” recovery in the source bin, transport and place the object into the central box without collisions or drops, and finally execute $S$ . The paper reports that ranking was determined by number of fully successful packing tasks and then by cumulative completion times; NeSyPack won First Prize in the WBCD competition at ICRA 2025 (Li et al., 6 Jun 2025).

A second documented instantiation is the WBCD Table Service Track. There, an ARX X7s dual-arm humanoid is evaluated on three sequential subtasks repeated as many times as possible in 30 minutes: tablecloth unfolding, food container opening/closing, and pizza placement with container packing. The scoring rule is

$\mathrm{Score}_{\text{subtask}} = (s / \beta) \times \alpha,$

with $s$ the base score, $\beta$ the completion time in seconds, and $\alpha$ the autonomy coefficient, where $0.5$ corresponds to in-person teleoperation, $O_i$ 0 to remote teleoperation, and $O_i$ 1 to autonomous execution. This structure makes WBCD a systems-level evaluation rather than a pure policy-learning benchmark: speed, reliability, precision, and autonomy all enter the objective simultaneously (Li et al., 20 Aug 2025).

Taken together, these reports define WBCD as a task-driven examination of dual-arm competence under competition constraints. A plausible implication is that WBCD should be understood less as a single benchmark with one canonical protocol than as a family of challenge settings organized around practical two-arm manipulation under real hardware limitations.

2. Manipulation repertoire documented in the literature

The most direct answer to what bimanual systems can do is provided by the task sets themselves. In tactile dual-arm manipulation, Bi-Touch introduces three contact-rich tasks tailored to tactile feedback: bi-pushing, bi-reorienting, and bi-gathering. Bi-pushing requires collaborative motion of a large planar object along a straight or sinusoidal 2D trajectory; bi-reorienting rotates a free object of length in $O_i$ 2 mm about its center while keeping it in place; bi-gathering drives two separate objects until their centers approach within $O_i$ 3. On real hardware, the reported bi-pushing error for a 400 mm object was $O_i$ 4 mm for a box, $O_i$ 5 mm for a tube, and $O_i$ 6 mm for a loudspeaker; real bi-reorienting on objects of size 70–193 mm achieved translation errors of 12.5–19.5 mm and orientation errors of 7.5–13.4°, while an emergent “brick gathering” experiment reached 95 % success in 20 trials (Lin et al., 2023).

A different capability profile appears in role-structured dexterous tasks. BUDS evaluates pepper grinding, jacket zipping, marker capping, and vegetable cutting on a UR16e dual-arm platform. Given only 20 demonstrations, BUDS achieves 76.9% task success across the task suite and 52.7% success on out-of-distribution objects within a class. The key empirical point is that these are not merely simultaneous reaches; they are tasks in which one arm fixtures part of the environment while the other performs a precision action, often with repeated restabilization (Grannen et al., 2023).

One-shot demonstration systems expand the documented task inventory further. VLBiMan is evaluated on ten real-robot tasks, including six atomic bimanual skills—plug-pen, pen-into-cup inserting, bottle-unscrew, pour-water, pump-press, and spoon reorient—two long-horizon compositions—reorient+unscrew and unscrew+pour—and two multi-stage tool-use tasks, scoop with spoon and funnel-assisted pouring. The framework is explicitly designed for category-level variation, scene changes, visual clutter, and mixed synchronous/asynchronous dual-arm usage (Zhou et al., 26 Sep 2025).

At the benchmark level, PerAct² extends RLBench to 13 bimanual tasks with 23 unique task variations, including Push Box, Lift Tray, Sweep Dustpan, Put Item in Drawer, Put Bottle in Fridge, Push Two Buttons, Pick Up a Plate, Lift a Ball, Straighten Rope, Handover an Item, Handover Item (Easy), Pick Up Notebook, and Take Tray Out of Oven. These tasks collectively cover heavy-object transport, cooperative stabilization, assembly-style alignment, non-prehensile handling, deformable interaction, and handover protocols (Grotz et al., 2024).

In-hand reconfiguration introduces a distinct class of capability. The dual-limit-surface framework models two cooperative frictional patch contacts on either side of an object and plans alternating sliding and sticking motions. The resulting behaviors include continuous in-hand regrasp without object release or external fixtures, precise pose control by actively choosing which palm sticks and which slides, and robust performance on inclines, varied object shapes, and under gravity (Dang et al., 2024).

3. Sensing, perception, and representation

WBCD-relevant systems span a wide range of sensing regimes, from local contact geometry to scene-level affordance grounding. In Bi-Touch, each end-effector carries a TacTip optical tactile sensor, and the control observation fuses two tactile streams with proprioception:

$O_i$ 7

Each TacTip camera captures $O_i$ 8 px depth images at approximately 40 Hz, and a conditional GAN translates real TacTip images into simulated depth maps to close the sensory domain gap. This representation is explicitly local, contact-centric, and deformation-aware (Lin et al., 2023).

ViTaMIn-B addresses the complementary problem of collecting large-scale bimanual, contact-rich demonstrations with reliable sensing and tracking. Its DuoTact sensor is a compliant visuo-tactile device with a flexible TPU frame, dual contact faces, an interior LED strip, and a $O_i$ 9 Hz RGB camera. Rather than using raw tactile images as policy input, ViTaMIn-B reconstructs global deformation as a 3D point cloud and downsamples it to 256 points; tactile features are encoded with a shared PointNet, visual features with a DINO-pretrained ViT-B/16, and proprioception with a 2-layer MLP. For bimanual pose capture, two Meta Quest 3 controllers provide 6-DoF poses at 72 Hz in a shared world frame. In the Weight Placement ablation, this produced a 100 % valid demonstration rate versus 16 % for SLAM, and policy success of 0.7 versus 0.1 (Li et al., 8 Nov 2025).

Object-centric relational representations form another major line. Bi-KVIL extracts Hybrid Master-Slave Relationships (HMSR) among objects and hands from a small number of human demonstration videos. Relations are expressed in local frames attached to master objects, with keypoints discovered from Dense Object Nets, RAFT tracking, segmentation, and depth reprojection. The resulting task representation is described as object-centric, embodiment-independent, and viewpoint-invariant, and it supports relation types such as point-to-point, point-to-line, point-to-plane, point-to-curve, point-to-surface, and pose constraints (Gao et al., 2024).

Scene-agnostic bimanual planning instead begins with visual affordance reasoning over an entire scene. The Visual Point Grounding module extracts object points $N_{O_i}$ 0 and interaction points $N_{O_i}$ 1, using Grounded-SAM, raycasting, affordance prompts, and Set-of-Marks descriptors. These are assembled into an adjacency graph that captures which objects lie within simultaneous reach of both hands, thereby turning two-arm coordination into a grounded scene-graph problem rather than a purely reactive control problem (Lee et al., 10 Dec 2025).

Unified visuomotor approaches attempt to bridge these extremes. CUBic consumes wrist-camera images, a top-view image, and proprioceptive states, then produces arm-specific latent tokens coordinated through two codebooks with a shared index. Its premise is that independence and coordination should emerge from a shared tokenized perceptual structure rather than from either full decoupling or hand-crafted cross-arm coupling (Wang et al., 13 May 2026).

4. Coordination strategies and learning architectures

A recurrent theme in WBCD-related work is that bimanual coordination is usually made tractable by structural decomposition. BUDS decomposes control into a stabilizing arm and an acting arm. The stabilizer selects a 3D keypoint $N_{O_i}$ 2 to hold fixed, using a keypoint model $N_{O_i}$ 3, while a restabilizing classifier $N_{O_i}$ 4 determines when the current fixture is no longer effective; the acting arm then executes a behavioral-cloning policy $N_{O_i}$ 5 in the resulting quasi-static environment. This reduces the effective control dimensionality of a 14-DoF dual-arm problem by converting it into alternating fixture selection and single-arm action (Grannen et al., 2023).

Bi-Touch uses deep reinforcement learning rather than imitation. Its policy is a deterministic Gaussian-parameterized actor $N_{O_i}$ 6 with twin-critic $N_{O_i}$ 7, trained with PPO for $N_{O_i}$ 8 environment steps per task. For the sparse bi-gathering problem, the paper introduces a Goal-Update Mechanism (GUM) that generates $N_{O_i}$ 9 subgoals on the line between objects and reassigns each arm every $S$ 0 steps; in the real-to-sim transfer adjustments, $S$ 1 and $S$ 2 are used, together with a 2-stage curriculum. The paper explicitly characterizes GUM as an implicit curriculum compatible with on-policy learning (Lin et al., 2023).

VLBiMan decomposes one-shot human demonstrations into invariant primitives, or “anchors,” and variable modules. Adaptation to a new scene begins with a semantic parser,

$S$ 3

followed by a geometric feasibility problem,

$S$ 4

The framework then composes invariant and adapted modules under kinematic and collision constraints. At low level, it uses progressive IK refinement, dynamic collision compensation, and mixed synchronous/asynchronous scheduling; the latter is reported to speed up tasks by $S$ 5 (Zhou et al., 26 Sep 2025).

Large-model methods introduce yet another decomposition principle. BiCICLe casts bimanual control as a leader-follower in-context learning problem. A Leader LLM first predicts a single-arm trajectory, then a Follower predicts the other arm’s trajectory conditioned on the Leader’s plan. This replaces a monolithic $S$ 6 action prediction with two sequential $S$ 7 predictions, and the framework extends to Arms’ Debate and an LLM-as-Judge selection stage without any fine-tuning (Palma et al., 22 Apr 2026).

At the systems level, NeSyPack combines hierarchical symbolic task reasoning, a symbolic skill graph, supervised perception modules, imitation-learned manipulation skills, and motion planning solved with the Convex Feasible Set algorithm. By contrast, CUBic couples perception and control through a unified diffusion policy conditioned on quantized dual-arm latent tokens. These two systems exemplify opposite ends of the design spectrum—neuro-symbolic modularity and end-to-end tokenized visuomotor generation—yet both are explicitly coordination-aware (Li et al., 6 Jun 2025, Wang et al., 13 May 2026).

5. Evaluation regimes and empirical performance

The empirical status of WBCD-related systems is best understood through the benchmarks and competition reports that quantify coordination, success, and failure across different task families.

Setting	Reported metric	Result
WBCD packing	Competition outcome	First Prize; > 90% overall success (Li et al., 6 Jun 2025)
WBCD Table Service Track	Competition outcome	8/9 fully scored; total score ≈133 (Li et al., 20 Aug 2025)
TWIN with BiCICLe + Best-of-N	Average success rate	71.1% (Palma et al., 22 Apr 2026)
RoboTwin with CUBic	Average success rate	51.8%; +12.0% absolute over the strongest baseline (Wang et al., 13 May 2026)
BiCoord single-task benchmark	Best reported single-task SR	Pi0: 46.4% SR, 62.6% SSR, TL 380 steps (Peng et al., 7 Apr 2026)
BiManiBench Tier 3	Low-level control SR	GPT-5: 66.8% (Wu et al., 9 Feb 2026)

Competition performance and benchmark performance measure related but non-identical notions of capability. In the WBCD packing task, NeSyPack reports 100% success for cube, large, and cuboid categories in competition trials, and 83.3% for sphere and stacked objects. In the Table Service Track, the MilkDragon system completed nine full rounds in 30 minutes with average per-round time of approximately 3 min 20 s, combining remote teleoperation and an ACT-based autonomous module trained from 100 demonstrations (Li et al., 6 Jun 2025, Li et al., 20 Aug 2025).

Simulation benchmarks expose the gap between isolated successes and general coordination competence. PerAct² reports average task success rates of 5.9 for ACT, 10.5 for RVT-LF, 17.5 for PerAct-LF, and 16.8 for PerAct² across 13 tasks, showing that many language-conditioned 6-DoF bimanual tasks remain brittle even under noiseless RGB-D simulation. The paper also reports that PerAct² converges in approximately 54 h versus 89 h for PerAct-LF and 231 h for RVT-LF on an NVIDIA A40 (Grotz et al., 2024).

BiCoord shifts attention from isolated task completion to explicit spatial-temporal coordination. Its 18 tasks have average trajectory length 361 timesteps and average number of stages 4.27, and it introduces Minimum Relative Distance (MRD), Average Relative Distance (ARD), Simultaneous Movement Time (SMT), Simultaneous Movement Percentage (SMP), and Spatial-Temporal Integral (STI). These metrics quantify proximity, simultaneity, and coupled spatial-temporal activity rather than only final success (Peng et al., 7 Apr 2026).

BiManiBench provides an analogous hierarchy for multimodal LLMs. It separates Tier 1 spatial reasoning, Tier 2 high-level planning, and Tier 3 low-level end-effector control, and uses a Gaussian-Weighted Spatial Score for left-right arm assignment in addition to success rate and failure categorization. The benchmark’s main finding is that high-level reasoning proficiency does not imply robust dual-arm grounding or control (Wu et al., 9 Feb 2026).

6. Persistent limitations and research directions

Despite the breadth of documented tasks, WBCD-related systems continue to exhibit recurrent failure modes. In Bi-Touch, the sim-to-real gap initially caused the learned bi-reorienting policy to squeeze objects too hard, risking sensor damage; the authors mitigated this by retuning simulated TacTip stiffness and damping, penalizing excessive contact depth, and exposing the policy to larger target angles. The same study identifies moving goals in bi-gathering as too sparse for learning without GUM, and notes that workspace limits under repeated perturbations may require dynamic workspace re-centering or adaptive subgoal placement. It also lists full 6-DoF in-hand regrasping, free-space lifting, and modeling shear and dynamic friction effects as remaining challenges (Lin et al., 2023).

Role-decomposed visual imitation also has clear limits. BUDS relies purely on vision, so occlusions and drastically novel appearances can break both the keypoint model and the restabilizing classifier; role assignments are fixed by task, and tactile or force feedback is absent. ViTaMIn-B, although it improves demonstration capture substantially, is still confined to tabletop scenarios, lacks a global top-down or third-person camera in the loop, and its 256-point edge representation does not directly encode local pressure distribution across the silicone (Grannen et al., 2023, Li et al., 8 Nov 2025).

Long-horizon and tightly coupled tasks remain especially difficult. BiCoord shows that representative policies fail when precision demands tighten below centimeter scale, task length extends beyond a few hundred timesteps, or inter-arm roles must invert mid-task while anticipating future trajectories. In Handover Block With Bowls, all benchmarked methods achieve near 0 % SR, indicating that millimeter-scale alignment and synchronized pouring remain outside the range of current generic policies (Peng et al., 7 Apr 2026).

The same diagnosis appears in large-model evaluation. BiManiBench reports that MLLMs frequently fail through state-estimation misjudgment, end-effector allocation error, action sequencing error, bimanual conflict or collision, and malformed action parameters. Scene-agnostic hierarchical planning mitigates some of these issues through affordance grounding and skill libraries, but it still faces LLM hallucinations in multi-step subgoal reasoning, source/target inversion in asymmetric tasks, and the limitations of rigid, open-loop execution; the paper explicitly points to closed-loop replanning, dynamic contacts, deformable objects, force-sensitive operations, and real-hardware transfer as future directions (Wu et al., 9 Feb 2026, Lee et al., 10 Dec 2025).

This suggests that the current frontier of WBCD is not the existence of isolated dual-arm successes, which are now well documented, but the unification of precision contact, long-horizon memory, disturbance recovery, adaptive role exchange, and closed-loop multimodal feedback in a single system. Under present evidence, bimanual robots can pack, serve, push, reorient, gather, hand over, unscrew, pour, scoop, wipe, fixture, and reconfigure in-hand; what remains unresolved is making these behaviors routine across unseen scenes, embodiments, and contact conditions without sacrificing coordination fidelity.