Robotic imitation learning has achieved impressive success in learning complex manipulation behaviors from demonstrations. However, many existing robot learning methods do not explicitly account for the physical symmetries of robotic systems, often resulting in asymmetric or inconsistent behaviors under symmetric observations. This limitation is particularly pronounced in dual-arm manipulation, where bilateral symmetry is inherent to both the robot morphology and the structure of many tasks. In this paper, we introduce EquiBim, a symmetry-equivariant policy learning framework for bimanual manipulation that enforces bilateral equivariance between observations and actions during training. Our approach formulates physical symmetry as a group action on both observation and action spaces, and imposes an equivariance constraint on policy predictions under symmetric transformations. The framework is model-agnostic and can be seamlessly integrated into a wide range of imitation learning pipelines with diverse observation modalities and action representations, including point cloud-based and image-based policies, as well as both end-effector-space and joint-space parameterizations. We evaluate EquiBim on RoboTwin, a dual-arm robotic platform with symmetric kinematics, and evaluate it across diverse observation and action configurations in simulation. We further validate the approach on a real-world dual-arm system. Across both simulation and physical experiments, our method consistently improves performance and robustness under distribution shifts. These results suggest that explicitly enforcing physical symmetry provides a simple yet effective inductive bias for bimanual robot learning.
Figures & tables
Fig. 1: Symmetry-equivariant policy learning for bimanual manipulation. A symmetry transformation S defines a left–right exchange of the scene in the image coordinate frame, generating a symmetrically equivalent task instance. Given an observation O and its transformed counterpart S(O) , the shared policy π is trained to produce equivariant predictions under the same transformation, i.e., π(S(O))≈S(π(O)) . The symmetry constraint is imposed at the action level and does not require architectural modifications.
Fig. 2: Overview of the proposed EquiBim framework. Given an observation O and its symmetrically transformed counterpart S(O) , where S denotes a left-right reflection defined in the image frame, both inputs are processed by a shared policy π . The policy takes visual observations together with proprioceptive states as input and produces action predictions π(O) and π(S(O)) . A symmetry-equivariant loss Lsym enforces the predicted actions to transform consistently under S during training.
Fig. 3: Visualization of eight simulated bimanual manipulation tasks used for evaluation. Subfigures (a)-(g) are adapted from RoboTwin [ 19 ] . For each task, the left image shows the initial state and the right image shows the goal state.
Observation + Action
Baseline
+ EquiBim
Δ
DP
Image + Joint
34.1
43.6
+9.5
Image + EE
37.3
40.0
+2.7
DP3
Point Cloud + Joint
73.5
77.9
+4.4
Point Cloud + EE
74.5
77.8
+3.3
TABLE I: Average success rate (%) across observation–action settings on RoboTwin.
Fig. 4: Overview of the real-world bimanual manipulation setup. Two LeRobot SO101 robotic arms operate in a shared workspace and are observed by a Logitech C920x camera. The system is evaluated on object handover and hook-hanging tasks under a top-down camera configuration.
Fig. 5: Real-world task executions on the bimanual LeRobot platform. The figure presents three tasks, namely Object Handover (Banana), Hook Hanging (Drumstick), and Hook Hanging (Toy Chicken), each depicted across three stages: initial, intermediate, and final. Each row shows the temporal progression of the manipulation process under a top-down RGB camera view.
Fig. 6: Generalization under distribution shift on the real-world bimanual platform. Rollouts are shown under the training distribution (left) and shifted conditions (right) for object handover, drumstick hook hanging, and toy chicken hook hanging. Each visualization overlays multiple rollouts (out of 10 total trials per setting) using transparency to illustrate trajectory variability. For clarity, only a subset of representative rollouts is displayed. The shifts include symmetric orientation changes, demonstration quality variation, and randomized initialization.
Method
Training Distribution
Shifted Distribution
Banana
Drumstick
Toy Chicken
Banana
Drumstick
Toy Chicken
ACT
3/10
3/10
8/10
0/10
1/10
4/10
ACT + EquiBim
6/10
3/10
9/10
5/10
4/10
6/10
TABLE III: Success rate in real-world experiments under training and shifted object distributions.