We approach pose retargeting from first-principles.
Clip 1/3: Wrist and Elbow Fast Motion
Clip 2/3: Fine Wrist Movements
Clip 3/3: Torso Control and Near Singularity Movements
We propose SEW-MIMIC:
an analytical method for exact human-robot bimanual pose retargeting
Iterative nonlinear optimization solvers are slow and may not converge, especially near singularities. Instead, we propose a constant-time, analytical approach that handles singularities with ease.
Clip: Real-time human (right-half screen) to robot (left-half screen) pose retargeting using SEW algorithm. Singularity Margin shows how hard pose retargeting is for iterative methods.
Our analytical approach enables a fast safety filter that empowers instead of restricts. Users can perform highly dynamic maneuvers without fear of self-collision.
Clip 1/3: Filtering Aggressive Self-Collision
Clip 2/3: Continuous Near Self-Collision Movement
Clip 3/3: Near Self-Collision Fine Movements
The SEW action representation can be used to train fine dexterous manipulation. We achieved a similar success rate on tasks introduced in DexMimicGen using their trajectories.
Clip 1/2: Coffee Making Task with GR1
Clip 2/2: Object Pouring Task with GR1
Our closed-form, CPU-only solution conserves computational resources that can be used for other, unstructured problems.
Clip 1/3: Making the Entire Arm Useful
Clip 2/3: Towards More Dynamic Demonstrations
Clip 3/3: Capable of Bimanual Coordination
Our method easily generalizes to other robot platforms, including bimanual manipulators with twice the arm span and different wrist structure.
Clip 1/2: Dual-Kinova3-14-DoF Robot Crossing Arms
Clip 2/2: Self-Collision Filter Enabled for the Arm Rolling Motion