Dexterous Hand Control: Admittance Control vs. Motion Retargeting
Sources: Research synthesis, 2026-05-05 Raw: Admittance Control vs Retargeting Updated: 2026-08-12
Overview
When using a motion capture glove to control a dexterous robot hand, two fundamentally different control architectures are available: fingertip Cartesian endpoint Admittance Control (Task Space) and Motion Retargeting (Joint/Geometric Space). The choice determines contact compliance behavior, intermediate joint configuration fidelity, and the type of data captured for downstream imitation learning. For LfD data collection, retargeting is preferred (full joint configuration data for behavior cloning). For contact-rich precision assembly, admittance control is safer and more natural.
The Core Difference
Fingertip Admittance Control operates in Cartesian Task Space: the glove outputs fingertip 3D positions → these become the desired endpoint positions for each finger → Admittance Controller maps contact forces to endpoint position corrections → IK solves for joint angles each control cycle.
Contact force F_contact
↓
Admittance: M·ẍ + B·ẋ = F_contact
↓
Endpoint correction Δx
↓
Desired endpoint = glove fingertip position + Δx
↓
IK → joint angles
Motion Retargeting operates in Joint/Geometric Space: the glove outputs joint angles → a mapping function converts them to robot joint angle commands. Three implementation levels:
- Direct angle mapping: human joint angle → scale/offset → robot joint angle (simplest, ignores morphology differences)
- FK-IK pipeline (e.g., DOGlove): FK computes human fingertip positions → IK finds robot joint angles with scaling factors
- Geometric optimization (e.g., ByteDexter Keyvector): extract geometric vectors from human hand → minimize human-robot geometric vector difference → robot joint angles
Key Differences
| Dimension | Fingertip Admittance Control | Motion Retargeting |
|---|---|---|
| Control space | Cartesian Task Space | Joint/Geometric Space |
| Contact compliance | Built-in (Admittance equation) | Requires separate force control layer |
| Intermediate joint fidelity | Low (IK Null-space decides) | High (mapping preserves configuration) |
| Morphology difference handling | IK naturally absorbs scale differences | Must be explicitly designed (scaling, joint coupling) |
| IK singularity risk | Present (fully extended/bent fingers) | None (joint space operation) |
| LfD data quality | Rich contact force information | Complete joint configuration information |
| Hardware requirement | F/T sensor per finger | Joint angle sensors only |
| Best task type | Contact-rich precision tasks | Configuration-critical dexterous tasks |
The Critical Fork: Contact Compliance
Admittance control has built-in contact compliance — when the finger contacts an object, the contact force drives a position correction, making the robot hand “yield” naturally. This is the correct behavior for grasping fragile objects or performing precision assembly.
Pure retargeting in position control mode is rigid — contact generates unconstrained force. You must add a separate impedance or force control layer to achieve compliance. This is an engineering addition, not inherent to the architecture.
Intermediate Joint Configuration
Admittance control guarantees that fingertips reach target positions; what the intermediate joints (PIP, DIP) do is determined by the IK Null-space. The robot’s intermediate joint configuration can look nothing like the human’s — only the fingertip position matches.
Retargeting preserves the visual similarity of the entire hand configuration. For behavior cloning, this matters: the training data contains full joint configuration information at every timestep, which is what BC algorithms need to predict joint-space actions.
LfD Data Quality Implications
Retargeting data (for LfD): complete joint configuration at every timestep → behavior cloning can directly learn to map observations to full joint-space actions. The natural choice for policy training.
Admittance control data (for LfD): rich contact force information → better for policies that must handle contact-rich tasks where force signals matter. The natural choice for assembly, insertion, and manipulation of fragile objects.
Practical Recommendation
Goal: LfD data collection for imitation learning → Retargeting first. Full joint configuration is more valuable for BC. Add a lightweight impedance layer for equipment protection.
Goal: contact-rich precision manipulation (assembly, insertion, fragile objects) → Admittance endpoint control. Natural compliance, intuitive operator experience. Must solve IK singularity issues (damped least squares is the standard).
Hybrid: some systems use retargeting as the primary mapping with a local contact impedance layer at the robot side — the practical compromise for general teleoperation.
Open Questions
- Can some fingers use admittance control (where contact compliance is critical) while others use retargeting (where configuration similarity is needed)? Mixed-mode control is unexplored.
- How should IK Null-space Secondary Objectives be designed to maximize intermediate joint configuration similarity to the human hand?
- Is there a minimum force sensing requirement (sensors per finger, resolution) below which admittance control degrades in practice?