Extending Object-Centric Dexterous Manipulation from Bi-Manual Arms to Humanoids

In this project, we extend Obj-Dex from dexterous hands to humanoid robots by retaining its hierarchical object-centric framework

Background

Explains the original idea behind the project.

Planning

Details the steps taken to organize the project.

Image showing the prototype being tested in a real-world setting.

Execution

Describes how the project was brought to life.

Outcome

Summarizes the results and impact achieved.

→

Obj-Dex is an object-centric manipulation framework that leverages human demonstration data and a hierarchical policy, combining a high-level trajectory planner with a low-level reinforcement learning controller to achieve robust in-hand and bimanual manipulation.

Extending the policy to humanoids

Inputs to the Policy
The policy receives detailed information about the hand, object, and task: Hand State: joint positions and velocities, fingertip poses, and actuator states. Object State: object position and orientation (quaternion), along with linear and angular velocities. Task Information: the goal pose of the object and task-specific features, such as a grasp indicator or success flag.

RL Algorithm: PPO Policy Network: Single shared actor-critic MLP Actor outputs. Gaussian distribution parameters (mean + learned/fixed log-std). Critic outputs scalar value estimate. Optimization Technique: On-policy training using PPO Loss components: policy loss (advantage-weighted log-prob), value function MSE loss, and entropy regularization. Gradient updates performed with PyTorch optimizer Rollouts: Collected across thousands of parallel Isaac Gym environments

Outputs of the Policy Continuous joint-level actions for the dexterous hand: Target joint positions OR joint torques depending on env config Actions executed in Isaac Gym to move the hand and manipulate the object
Environment returns rewards + next state for the next training iteration