UniLatent: Cross-Embodiment Transfer via Latent Observation Alignment
UniLatent: Cross-Embodiment Transfer via Latent Observation Alignment
This paper introduces UniLatent, a framework for cross-embodiment robot policy transfer that aligns source and target robot observations in a shared latent space without pixel-level image translation. The method achieves up to 93% zero-shot success across three simulation tasks and two target robot arms, outperforming the strongest pixel-translation baseline by over 30 points on average. In real-world xArm-to-Franka transfer, fine-tuning with only 10 target demonstrations per task yields 27/30 successful rollouts. Code, videos, and data are available at https://selina2023.github.io/unilatent/.
The authors present UniLatent, a framework for transferring robot policies across different embodiments by aligning observations in a shared latent space, achieving up to 93% zero-shot success in simulation and 27/30 successful rollouts in real-world xArm-to-Franka transfer with minimal fine-tuning.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S3
Reported numbers
target demonstrations per task
10
View original evidence
real-world xArm-to-Franka transfer, fine-tuning with only 10 target demonstrations per task yields 27/30 successful rollouts. Code, videos and data are available at https://selina2023.github.io/unilatent/.
Open source S4simulation tasks
3
View original evidence
t robot observations in simulation and aligns a source visual encoder to a frozen target encoder. A policy trained on encoded source demonstrations can then run directly on target observations, with no pixel-level image translation. Across three simulation tasks and two target robot arms, UniLatent reaches up to 93% zero-shot success, exceeding the strongest pixel-translation baseline by over 30 points on average. In
Open source S3
- code: Available Artifact link Open source S4
- dataset: not_reported
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
t robot observations in simulation and aligns a source visual encoder to a frozen target encoder. A policy trained on encoded source demonstrations can then run directly on target observations, with no pixel-level image translation. Across three simulation tasks and two target robot arms, UniLatent reaches up to 93% zero-shot success, exceeding the strongest pixel-translation baseline by over 30 points on average. In
Open source S3
real-world xArm-to-Franka transfer, fine-tuning with only 10 target demonstrations per task yields 27/30 successful rollouts. Code, videos and data are available at https://selina2023.github.io/unilatent/.
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org