← AI Techniques Radar
Inference efficiency

Unified inertial encoder via kinematic IMU synthesis for cross-body motion transfer

Create a shared inertial input space for both human and humanoid motion by synthesizing kinematic IMU signals at corresponding body sites, then train a single continuous-time encoder to jointly predict velocity and align motion with text within that shared space.

Discovered 2026-11-10 · Sources checked 2026-11-10

This summary was generated with AI assistance.

Why it matters

It reduces the mismatch between human and humanoid sensing and body geometry, enabling models trained on human motion to be used for humanoid motion estimation and language-grounded motion understanding without requiring simulation or per-robot fine-tuning.

What changed

Instead of using robot-specific IMU layouts and separate encoders, the approach builds a unified inertial interface by mapping motions from different bodies into corresponding “kinematic IMU” representations, then performs joint training for velocity estimation and motion–text alignment on both human data and retargeted humanoid motion.

Where it helps

  • Human-to-humanoid inertial odometry transfer
  • Language-grounded motion understanding from motion-text pairs
  • Zero-shot deployment of inertial motion models across robot platforms
  • Cross-body generalization for sensor-based motion retrieval

Evidence & reproduction

Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.

Limitations & open questions

The method relies on having kinematic/pose data sufficient to perform IMU synthesis and retargeting, and it targets transfer through the shared inertial representation (so performance depends on how well the synthesized IMU signals match the target robot’s true sensing).

Related World Models papers

Sources