Bimanual Manipulation
How it works
1) Task classification: per Smith et al. (2012), tasks split into uncoordinated (arms perform independent actions) and coordinated, and the latter into symmetric (both hands play an equal role) and asymmetric (a stabilizing hand plus an acting hand). 2) Coordination: a controller or policy computes both arms' motions so they are temporally and spatially consistent; when both grasp one object a closed kinematic chain forms, so one must control not only position but also internal forces (squeezing/stretching the object). 3) Perception: multi-camera vision (often wrist and external cameras), and increasingly touch (tactile sensors), supplies the state of both arms and the object. 4) Control/policy: classically โ constrained motion planning and force-position control; today โ learned visuomotor policies (ACT, Diffusion Policy, BiKC, VLA models) trained on dual-arm teleoperation demonstrations (e.g. ALOHA). 5) Execution: the policy produces coordinated trajectories/actions for both arms, often with inter-arm collision avoidance and load sharing.
Problem solved
A single arm is insufficient for a vast class of everyday and industrial tasks: holding a large or deformable object, assembly that requires simultaneous stabilizing and acting, opening containers, threading, folding fabric, or handing an object off "from hand to hand". Bimanual manipulation addresses tasks that are physically impossible with one hand or that require maintaining a closed kinematic chain and controlling internal forces between two grippers.
Components
Computing both arms' motions so they are temporally and spatially consistent and realize a shared goal.
When both arms grasp the same object they form a closed chain; one arm's motion imposes constraints on the other.
Managing squeezing/stretching forces on a jointly held object and distributing load between the arms.
Distinguishing a stabilizing hand from an acting hand (asymmetric coordination) or equal roles (symmetric) โ central to the Krebs & Asfour taxonomy.
Ensuring two arms working in a shared workspace do not collide with each other or the object.
Aligning the timing and motion phases of both arms, crucial for tightly coupled tasks.
Implementation
With a shared grasp, two arms can squeeze or stretch the object; without internal-force control it is easy to damage or drop it.
The arms share a workspace and can collide, especially in tight configurations or with learned policies.
Temporal or phase drift between the arms breaks tightly coupled tasks (e.g. hand-over).
The doubled action space raises demonstration-data requirements and makes learning stable policies harder.
Two arms require mutual calibration; errors in inter-frame transforms lead to imprecise coordination.
Evolution
Smith et al. systematize the field, introducing the split into uncoordinated and coordinated (symmetric/asymmetric) tasks.
Krebs and Asfour (IEEE RA-L) propose a taxonomy of human bimanual coordination based on hand roles, symmetry, and interaction.
Zhao et al. (RSS 2023) demonstrate fine-grained bimanual manipulation learned from a few minutes of demonstrations on low-cost hardware โ a turning point for learning-based bimanual.
Mobile ALOHA adds a mobile base and whole-body control; PerAct2 and RoboTwin provide benchmarks for learning bimanual manipulation; policies like BiKC advance the methods.
Hyperparameters (configurable axes)
Uncoordinated vs. coordinated โ whether the arms perform independent actions or cooperate toward a shared goal.
Within coordinated tasks: symmetric (equal roles) vs. asymmetric (stabilizing hand + acting hand).
Whether the arms hold a shared object (closed kinematic chain) or act on separate objects.
The combined degrees of freedom of both arms (plus grippers) โ roughly double that of a single-arm setup.