本文由 AI 分析生成
建立時間： 2026-05-30 來源： https://arxiv.org/abs/2507.03227
Summary
ByteDance Seed Robotics presents a Keyvector-based retargeting framework for dexterous teleoperation of the 20-DoF ByteDexter hand using a Manus Glove (25 DOF, 120 Hz). The method represents human-robot hand correspondences as 3D vectors between keypoints (fingertip to MCP), then solves a constrained optimization to minimize vector differences while handling kinematic constraints unique to the robot hand.
ByteDance Seed Robotics 提出以幾何向量（Keyvector）為基礎的靈巧手遙操作 Retargeting 方法。從 Manus Glove 提取 15 個向量後，透過約束優化求解關節角度，並以 400x 加權懲罰項強化捏取相關運動的保真度，處理了人機形態差異問題。
Prerequisites
- Keyvector — 3D vectors between hand keypoints (e.g., fingertip to MCP), invariant to overall hand scale
- Retargeting — mapping human hand motion to a morphologically different robot hand
- PIP-DIP coupling — underactuated constraint where proximal and distal interphalangeal joints are mechanically linked
Core Idea
Represent hand configuration as a set of 15 geometric vectors between keypoints rather than joint angles. Solve for robot joint angles that minimize the difference between human and robot Keyvectors, with scaling factors to compensate for size differences. Special handling for the non-anthropomorphic ByteDexter palm (motors integrated in palm) and PIP-DIP coupling via chain rule differentiation.
Results
| Metric | Value |
|---|---|
| Tracking frequency | 120 Hz (Manus Glove input) |
| Robot DOF | 20 (ByteDexter) |
| Pinch weight factor | 400x priority for grasp-critical vectors |
Limitations
- Author-stated: Thumb lacks true opposition (saddle joint difference), limiting manipulable object geometries
- Author-stated: Operator must continuously maintain low-level hand stability — high cognitive load
- Author-stated: No autonomous force regulation; operator controls contact force manually
- Unstated: No evaluation against alternative retargeting methods (MANO-based, etc.)
Reproducibility
- Code: Not open-sourced (ByteDance proprietary)
- Dataset: N/A
- Compute: Standard robot control loop
Insights
The 400x pinch weighting is a strong engineering choice: by dramatically prioritizing grasp-relevant vectors over non-grasp vectors, the optimization naturally produces retargeting that preserves object-contact quality even when the overall finger configuration differs. This is the right inductive bias for manipulation tasks.
The PIP-DIP underactuated constraint is a common challenge in dexterous hand retargeting that many papers gloss over. Handling it explicitly via chain rule differentiation in the optimization is technically sound and necessary for real deployment.
Connections
Raw Excerpt
“We calculate 15 Keyvectors and solve for joint angles that minimize the difference between robot and human Keyvectors, with collision avoidance and pinch-weighted penalties up to 400x.”