ROBOT POLICY REGISTRY
A curated starting point for discovering open robot policies and foundation models. We highlight projects with public technical resources that may inform future RobotReplica evaluations.
Inclusion does not indicate endorsement, hardware compatibility, or evaluation by RobotReplica. Verified performance remains on the site leaderboards.
FEATURED POLICIES
8 public projects selected for their relevance to generalist and cross-embodiment robot learning.
Robot foundation model
Xiaomi Robotics
A vision-language-action foundation model pretrained on more than 100,000 hours of embodiment-free UMI manipulation trajectories, then aligned with cross-embodiment robot data for mobile manipulation and efficient task adaptation.
Robotic manipulation foundation model
Qwen Team
A Qwen3.5-4B-based VLA with a flow-matching diffusion-transformer action expert, trained on approximately 38,100 hours of open robot data, egocentric video, and synthesized demonstrations across multiple embodiments.
Action reasoning foundation model
Ai2
An open 5B robot-control model that pairs the Molmo 2-ER embodied-reasoning backbone with a flow-matching continuous action expert, with released weights, training data, code, and LeRobot integration.
Generalist robot foundation model
NVIDIA
An open VLA for generalized robot skills, with multimodal inputs, a diffusion-transformer action head, and adaptation workflows for multiple embodiments.
Vision-language-action model
Physical Intelligence
An open-world generalist VLA trained across multiple robots and heterogeneous data sources, released through the OpenPI codebase with base checkpoints and fine-tuning examples.
Compact vision-language-action model
Hugging Face · LeRobot
A compact 450M-parameter VLA trained on community-contributed LeRobot datasets and designed to run and fine-tune on accessible hardware.
Cross-embodiment vision-language-action model
Tsinghua AIR
A scalable VLA that uses embodiment-specific soft prompts to share a transformer policy across heterogeneous robot platforms.
Open-source vision-language-action model
Stanford · UC Berkeley · TRI
A 7B-parameter open VLA pretrained on 970k robot episodes from Open X-Embodiment, with released checkpoints and fine-tuning support.
KEEP THE REGISTRY CURRENT
Send the project page, paper, or public checkpoint to the RobotReplica team. We review entries for relevance and source quality before adding them.
Suggest a policy