PA-BiCoop: A Primary-Auxiliary Cooperative Framework for General Bimanual Manipulation

Abstract: Bimanual manipulation is essential for advanced robotic systems because it offers higher efficiency and flexibility compared to single-arm configurations. However, existing approaches either lack inter-arm interaction or ignore the need for a dynamic division of labor, treating the arms as functionally equivalent. To address these limitations, this paper draws inspiration from human bimanual manipulation where one arm handles core operations and the other provides auxiliary support, and proposes PA-BiCoop, a new single-model bimanual cooperation framework with dynamic primary-auxiliary arm differentiation. PA-BiCoop categorizes robotic arms into primary and auxiliary arms with adaptively adjustable roles across task stages, employs two specialized decoders that share a global feature encoder: the primary decoder generates the primary arm's base-coordinate pose and core-task affordance heatmaps, and the auxiliary decoder outputs the auxiliary arm's relative pose in the primary arm's coordinate system. Moreover, we design a dynamic role assignment module to automatically map roles to left/right arms without manual pre-definition. This design facilitates inter-arm knowledge sharing and coordinated manipulation. Extensive experiments demonstrate that our PA-BiCoop achieves superior performance: it outperforms state-of-the-art baselines by 48% on average in RLBench2 simulation tasks and by over 50% on average in real world tasks, thereby verifying its effectiveness and advancement in bimanual manipulation.

• The paper introduces a single shared model that dynamically assigns primary and auxiliary roles, improving average success to 61.6–68.8% in RLBench2 and 85% on real-world tasks.
• A relative-coordinate auxiliary decoder, specialized asymmetric heads, and learned role switching enable stronger coordination, with ablations showing average drops of 24, 16, and 13.3 percentage points when key components are removed.
• The framework reduces sample demands and outperforms baselines substantially, but it still depends on annotated role labels and may struggle with long-horizon tasks or movements lacking stable inter-arm geometry.

PA-BiCoop addresses a structural deficiency in learning-based bimanual manipulation: existing frameworks either duplicate models per arm, sacrificing inter-arm knowledge sharing, or share a single model while treating both arms as functionally equivalent. The authors propose a single-model architecture that explicitly differentiates the two arms into dynamically assigned primary and auxiliary roles, motivated by the observation that human bimanual behavior is asymmetric and stage-dependent. The framework is evaluated on RLBench2 and on a physical dual-arm platform, reporting average improvements of 48% in simulation and over 50% in real-world tasks relative to state-of-the-art baselines (2606.28192).

Motivation and positioning

The paper organizes prior work into two paradigms. Dual-model approaches (RVT-LF, PerAct-LF, AnyBimanual, VoxAct-B, BUDS) assign each arm a dedicated model or decoupled stabilizing/acting roles; they inherit single-arm modeling advances but lack interactive information transfer and double model complexity. Single-model approaches (ACT, PerAct2, Kstar Diffuser, InterACT, YOTO) share representations across arms but either output both action spaces simultaneously without role distinction or rely on handcrafted coordination sequences. The central claim is that this role-agnostic treatment of functionally equivalent arms fails to capture the inherent asymmetry of bimanual tasks—e.g., one arm aligning a component while the other stabilizes it—and that explicit, dynamic role specialization is the missing ingredient.

Architecture

PA-BiCoop uses an RVT backbone as a shared global encoder over multi-view RGB-D images (128×128), language instructions, and proprioception, producing image and language tokens consumed by three heads:

• Primary decoder: eight transformer self-attention layers followed by convolutional layers with bilinear upsampling generate multi-view affordance heatmaps $H^P_T$ whose peaks give the primary arm's translation in the base coordinate system $C_{bc}$; joint features (heatmap-weighted spatial sum plus max-pooled features) feed an MLP predicting discretized Euler angles ($5^\circ$ bins) and binary variables (gripper state, collision-avoidance flag).
• Auxiliary decoder: one cross-attention layer and six self-attention layers predict the auxiliary arm's pose in the primary arm's coordinate frame $C_{pc}$, conditioned on a token distilled from the primary decoder's argmax outputs concatenated with a learnable query embedding. Because action tokens are far smaller than image/language tokens, cross-attention focuses on salient regions at reduced computational cost.
• Role assignment module: a learnable query $\xi$ attends to image/language tokens and then to auxiliary-decoder action tokens through six cross-attention blocks, outputting a binary variable mapping primary/auxiliary roles onto left/right arms without manual pre-definition.

A coordinate transformation converts the auxiliary prediction from $C_{pc}$ back to $C_{bc}$ via rotation-matrix composition for orientation and vector addition for translation. Training uses cross-entropy losses on heatmaps, rotation distributions, and binary variables for the primary arm; MSE on translation, cross-entropy on binaries, and a circular MSE loss on periodic Euler angles (with scaling factor $\kappa = 1/360$) for the auxiliary arm; and cross-entropy on $\xi$. Key-frame extraction restricts prediction to next key-frames, following PerAct/RVT conventions.

Two design choices carry most of the method's weight. First, predicting the auxiliary pose in $C_{pc}$ makes it invariant to manipulated-object displacement and reduces the auxiliary arm's absolute spatial-perception burden—an assumption that holds when inter-arm relative geometry is stable during coordination. Second, dynamic role switching ($C_{bc}$0) is supervised by an annotation schema added to the training data, meaning ground-truth role labels must be available at training time; the paper does not evaluate performance when such annotations are noisy or absent.

Simulation results

On ten RLBench2 tasks spanning synchronous, asynchronous, symmetric, and asymmetric scenarios, trained with 20 or 100 demonstrations and evaluated over 25 test episodes per task, PA-BiCoop reaches 61.6% average success with 20 demos and 68.8% with 100 demos, versus 11.6%/20% for AnyBimanual, 4%/14% for PerAct2, and 6%/6% for ACT. The strongest results appear in synchronous symmetric tasks: push box (84–88%), lift ball (100%), lift tray (80–88%), and sweep dustpan (96–100%). On long-horizon tasks (put in bridge/fridge, take out tray), success rises from under 10% for baselines to 45–68%, which the authors attribute directly to observation-driven role switching. Notably, Kstar Diffuser reports strong numbers on its subset of tasks (e.g., 83–98% on lift ball/push box at 100 demos) but was not evaluated on the full suite because code is unavailable, so the comparison against it is partial and taken from the original publication—a caveat the authors acknowledge.

Ablations

Ablations on three representative tasks isolate each component's contribution:

Configuration	Push Box	Put in Drawer	Take out Tray	Avg.
Full (P-A decoders, $C_{bc}$1, $C_{bc}$2)	88	60	68	72
w/o $C_{bc}$3	88	60	28	58.7
w/o $C_{bc}$4	44	32	68	48
w/o both	44	32	28	34.7
Primary-primary (no specialization)	20	24	12	18.7

Removing role assignment costs 13.3 points on average and disproportionately harms the long-horizon take out tray task (−40 points), confirming that fixed primary roles degrade precision over extended sequences. Predicting the auxiliary pose in $C_{bc}$5 instead of $C_{bc}$6 costs 24 points, supporting the relative-coordinate hypothesis. Replacing the auxiliary decoder with a second primary decoder (primary-primary) collapses performance by more than 16 points even below the no-$C_{bc}$7/no-$C_{bc}$8 variant, indicating that asymmetric decoding—not merely asymmetric coordinates—is essential.

Real-world results

On a dual Yahboom DoFbot system with a front-facing RGB-D camera, 15 demonstrations per task, and 10 evaluation trials, PA-BiCoop achieves 85% average success (90% grasp banana, 80% handover) versus 35% for PerAct2 and 5% for ACT. The >50-point margin despite very small demonstration counts suggests the architectural priors (relative-coordinate auxiliary prediction, role supervision) reduce sample complexity, though the task set is limited to two contact-rich but structurally simple scenarios on low-cost hardware.

Limitations and open questions

The authors concede that PA-BiCoop struggles with extremely long-horizon tasks involving dozens of sequential operations or extended pauses, since the current role assignment module lacks task-stage awareness or memory. Two further constraints are implicit in the design: the method requires ground-truth role annotations ($C_{bc}$9) during training, leaving open how roles could be learned self-supervised from unannotated demonstrations; and the auxiliary decoder's advantage rests on the assumption of stable inter-arm relative geometry, which may not hold for tasks with large independent arm excursions. Whether the reported gains transfer beyond RLBench2-style tabletop tasks and low-cost manipulators remains untested.

Conclusion

PA-BiCoop demonstrates that introducing explicit, dynamically assigned primary-auxiliary differentiation within a single shared model yields substantial gains in bimanual manipulation—48% average improvement on RLBench2 and over 50% in real-world trials—with ablations attributing the gains jointly to relative-coordinate auxiliary prediction, specialized decoding, and learned role switching. The main open questions concern scaling role assignment to long-horizon, multi-stage tasks and removing the dependence on manually annotated role labels.