Gasgoo Munich-Anyverse Dynamics has officially released and open-sourced the MWA™-Phoenix latent space world model. This edge-side, 7-billion-parameter fully pre-trained model secured leading results across six benchmarks—including LIBERO-Plus, LIBERO-Pro, RoboCasa GR1, Bench2Dex, RoboTwin 2.0, and RoboDojo—validating its capabilities in handling environmental disturbances, task generalization, complex manipulation, and real-world execution.
Image Source: Anyverse Dynamics
Across specific benchmarks, MWA™-Phoenix relies solely on RGB input, bypassing depth sensors. It claimed the top spot in LIBERO-Plus with a 94.63% overall success rate and led LIBERO-Pro at 66.48%. In the RoboCasa GR1 Tabletop evaluation, the model achieved an 82.67% average task success rate—marking an improvement of 7.50 percentage points over its predecessor.
Regarding cross-embodiment and collaborative operations, MWA™-Phoenix posted a 45.6% average success rate in the Bench2Dex evaluation and 72.42% in the RoboTwin 2.0 dual-arm manipulation test. Furthermore, the model delivered strong results across three robot embodiments and 18 tasks in the RoboDojo real-world leaderboard, validating its physical execution capabilities.
To enhance fine manipulation in real-world scenarios, Anyverse Dynamics simultaneously released Phoenix-ILA. This module co-encodes synchronous motion with tactile sequences, allowing direct contact feedback to inform action learning. In practical tasks, integrating this module boosted success rates for cucumber peeling and plate wiping from 20% and 75% to 40% and 90%, respectively.
The company also introduced Phoenix-PhysCanvas, an interactive simulation world model designed to provide a playable environment for reinforcement learning. Following post-training within this framework, average real-world success rates for test-tube and block manipulation tasks climbed from 58.33% and 80.00% to 75.00% and 86.67%, respectively.
Architecturally, MWA™-Phoenix employs DINOv3 as its visual backbone, fusing geometric priors with temporal information through dual-teacher distillation, while integrating environmental dynamic prediction and action generation into a joint learning framework. To date, Anyverse Dynamics has accumulated over 200,000 hours of multimodal data and is conducting larger-scale model training to support real-world dexterous manipulation tasks across commercial spaces and industrial production lines.
