Where to Touch, How to Contact: Hierarchical RL–MPC Framework for Geometry-Aware Long-Horizon Dexterous Manipulation

Zhixian Xie1,   Yu Xiang2,   Michael Posa3,   Wanxin Jin1
1Arizona State University 2University of Texas at Dallas 3University of Pennsylvania
Preprint Code Repo (Coming Soon) Video demo (Coming Soon)

Overview: we propose a hierarchical RL–MPC framework: the reinforcement learning (RL) performs geometric–kinematic reasoning at a higher level, while offloading the complex dynamic contact planning to contact-implicit model predictive control (MPC) at a lower level. The two levels are linked by a contact intention—a novel object-centric interface that specifies a surface contact location (geometry) together with a post-contact object-level subgoal (kinematics).


Comprehensive Video Demo:
real-world execution



How does it work?


Description of the image

Our proposed framework consists of (1) a higher-level policy for object-centric geometric and kinematic reasoning and (2) a lower-level policy for dynamic contact planning. The higher-level policy, trained using RL, predicts contact intention, which parameterizes the lower-level policy. The lower-level policy seeks to realize the contact intention through local contact-implicit model predictive control (MPC). The output of the lower-level policy is the robot action command.


Real-World Results


Geometry-generalized Pushing Task

In this task, the goal is to learn a unified policy that pushes a diverse set of letter-shaped objects from arbitrary initial poses to arbitrary planar target pose (position and orientation). 12 English letters are used in total: 6 letters (E, H, L, N, T, X) are used for training, and the remaining 6 (F, I, K, V, Y, Z) are held out for evaluation.


E_acc
F_acc
H_acc
I_acc
K_acc
L_acc
N_acc
T_acc
V_acc
X_acc
Y_acc
Z_acc

3D Object Reorientation Task

In this task, the goal is to reorient a cube from a randomized initial pose to a randomized 3D target pose.


cube_1_hd
cube_2_hd
cube_3_hd
cube_4_hd

More Simulation Results


Geometry-generalized Pushing Task


Push seen 1
Push seen 2
Push seen 3
Push seen 4
Push seen 5
Push seen 6
Push unseen 1
Push unseen 2
Push unseen 3
Push unseen 4
Push unseen 5
Push unseen 6

3D Object Reorientation Task


Cube 1
Cube 2
Cube 3
Cube 4


BibTeX

@article{TBD,
  author    = {Xie, Zhixian and Xiang, Yu and Michael Posa and Jin, Wanxin},
  title     = {Where to Touch, How to Contact: A Hierarchical RL–MPC Framework for Geometry-Aware Long-Horizon Dexterous Manipulation},
  journal   = {tbd},
  year      = {2026},
}