Dexterous Bimanual and Humanoid Manipulation

Team: Ankit Aggarwal, Parth Gupta, Swastik Mahaptra, Joshua Pen, Deepam Ameria, Shreya Shri Ragi, Srishti Gupta
Role: Robotics Engineer

This research investigates the challenge of enabling robotic systems to match human-like adaptability in unstructured environments. We developed a hierarchical, vision-driven control framework that decomposes complex manipulation into a high-level wrist planner and a low-level reinforcement learning (RL) controller. A primary focus of this work was evaluating the “embodiment gap”—testing whether object-centric policies trained on fixed-base arms can generalize to humanoid platforms like the Unitree G1.

Hierarchical Framework Overview

The system leverages an object-centric formulation, reasoning in task space rather than specific joint space to theoretically enable “write once, run anywhere” capabilities.

1. High-Level Wrist Planner: A Transformer-based generative policy that predicts 6-DOF wrist trajectories directly from monocular RGB image observations.
2. Low-Level Finger Controller: An RL policy trained in Isaac Gym that translates reference trajectories into joint-level actions for stable object interaction.

Key Contributions & Extensions

Transformer-Based Decoder

We replaced the standard MLP-based decoders in the baseline ArcticNet architecture with a Transformer-based decoder. This design introduces learnable query embeddings and multi-head attention, allowing the model to reason over structured hand-object relationships rather than treating parameter regression as an unstructured mapping.

Embodiment Transfer to Unitree G1

To bridge the gap between high-DOF Shadow Hands and the underactuated Inspire Hands of the Unitree G1, we implemented an explicit geometric mapping.

• Geometric Retargeting: Forward kinematics were used to compute embodiment-independent fingertip positions in Cartesian coordinates.
• Kinematic Mapping: We solved for finger flexion angles (alpha) and thumb rotation (beta) to respect the reduced kinematic complexity of the humanoid platform.

PPO Optimization & Scalability

We utilized Proximal Policy Optimization (PPO) with N=2048 parallel environments to stabilize high-variance gradient estimates. Our optimized configuration achieved a completion rate of 87.77%, exceeding the original Obj-Dex benchmark of 86.1%.

Results & Analysis

Our experiments demonstrated that while the high-level object-centric representation is transferable, the low-level execution requires embodiment-aware adaptation.

Quantitative Comparison (Transformer vs. Baseline)

Ablation Study: PPO Stability

The study confirmed that adaptive exploration enabled by learnable variance (Sigma) is critical for fine-grained control.

Technical Stack

• Simulators: Isaac Gym (parallelized rollouts), MuJoCo (motion fidelity), and Isaac Lab.
• Architectures: ResNet-152 backbone, Transformer decoders, and shared MLP [1024, 512, 256] Actor-Critic networks.
• Dataset: ARCTIC bimanual hand-object interaction dataset.

Video Presentation