Boundless Dynamics Open-Sources MWA™-Phoenix World Model, with Qianjue Tactile Sensors Supporting a Sweep of Six Leaderboards

Recently, Anyverse Dynamics released and open-sourced the MWA™-Phoenix latent-space world model base, a fully pretrained 7B-parameter on-device model. According to its published results, the model ranks first in six simulation and real-robot evaluations, including LIBERO-Plus (94.63% overall success rate), LIBERO-Pro, and the RoboDojo real-robot leaderboard. All real-robot manipulation tasks released this time use Qianjue tactile sensors as the tactile information input source. In this scheme, real contact information directly enters action learning and is also used to predict tactile feedback after actions. For Qianjue, this is also an application practice stemming from its long-term investment in tactile perception and data quality. Models understanding touch and using it for manipulation is becoming the intersection of the two parties' technical capabilities. Tactile data directly participates in action learning. Qianjue has long focused on contact information during manipulation. Only when tactile details correspond one-to-one with motion can the model learn from demonstrations how a manipulation is completed. In Anyverse Dynamics' research, Phoenix-ILA encodes synchronized motion and tactile sequences into the same interaction latent space. The model jointly learns motion, local tactile deformation, and force-related signals, preserving their temporal relationships through joint reconstruction. For example, when a peeler moves along a cucumber, how the blade conforms to the surface is also recorded; during plate wiping, tool movement and contact changes together constitute manipulation samples. Ablation experiments published by Anyverse Dynamics show that after adding Phoenix-ILA, the cucumber peeling success rate increased from 20% to 40%, and the plate wiping success rate increased from 75% to 90%. In manipulations that require continuous surface contact, real contact-process data can provide a direct basis for action learning. How the world model predicts tactile feedback. In addition to participating in action representation, tactile data is also used for policy training before deployment. Anyverse Dynamics' self-developed Phoenix-PhysCanvas is an action-conditioned visuotactile world model. After the policy outputs an action, it predicts the corresponding visual and tactile feedback, provides observations for subsequent manipulation, and supports GRPO reinforcement learning post-training. Training is conducted in an interactive simulation environment, and the results are then verified on real robots. Anyverse Dynamics first uses task demonstrations for supervised fine-tuning, then optimizes the policy in Phoenix-PhysCanvas. The average real-robot success rate for the test tube task increased from 58.33% to 75.00%; the block task increased from 80.00% to 86.67%, gains of 16.67 and 6.67 percentage points, respectively. What is presented here is the overall benefit of the visuotactile world model and reinforcement learning post-training. For tactile research, what is noteworthy is that the model's learning objective has been extended: in addition to using current contact information, it must also predict the feedback that may occur after an action is executed. Tactile feedback prediction likewise depends on real samples: the model must establish the correspondence between actions and contact changes from data, and the accuracy of this relationship is ultimately determined by tactile data quality. The tactile data standards behind real-robot tasks. All real-robot demonstrations released this time by Anyverse Dynamics use Qianjue photonic tactile sensors to acquire tactile signals. The related tasks include fine manipulations that require real-time perception of contact changes, such as unscrewing bottle caps, peeling, and wiping, as well as long-horizon tasks such as pencil case packing, sandwich making, and coffee scenarios. These real contact messages at different stages and of different types are all provided by Qianjue's tactile perception capabilities. In model training, the industry often focuses on whether the tactile modality has been introduced, but tends to overlook another question: can the acquired tactile signals be directly used as effective samples in training? In practical engineering, cross-unit consistency, spatiotemporal consistency, and physical consistency of tactile data are often underestimated, yet these indicators directly determine whether data can be reused at scale. Qianjue's work on sensors and embodied data systems is centered on precisely these points. Cross-unit consistency and repeat stability directly affect whether data collected by multiple machines can be aggregated for training. Embodied data collection usually requires multiple devices to work in parallel, accompanied by high-frequency repeated contact. If the responses of different sensors deviate, or if a single device exhibits signal drift after long-term use, the aggregated data will be mixed with noise. In sensor R&D and production, Qianjue has established testing standards for cross-unit consistency and repeat stability, ensuring that feedback from different devices under the same conditions is comparable and that the same device still outputs stably after repeated contact. Spatiotemporal consistency across modalities is the prerequisite for the model to establish causal relationships between actions and contact. Manipulation learning requires corresponding visual frames, motion trajectories, and tactile changes. Once tactile signals are misaligned with actions or images in frame rate or timestamp offset, it becomes difficult for the model to establish accurate interaction relationships. Qianjue carries out work around multimodal collection and data specifications, strictly aligning the temporal relationships among visual, tactile, and action signals. The accompanying real-time monitoring and process playback mechanisms support researchers in checking quality immediately during collection and reviewing samples before training. The physical consistency of data representation concerns whether tactile signals can truly reproduce the contact mechanics process. The contact signals recorded by sensors need to accurately correspond to contact deformation and force changes, and maintain a unified physical representation between real-robot collection and simulation training. Through tactile simulation and model research, Qianjue continuously verifies the physical consistency of data, ensuring that the recorded contact information conforms to real physical laws. When cross-unit consistency, spatiotemporal consistency, and physical consistency are all ensured, complete tactile signals can be transformed into training samples that models can directly use, becoming the foundational data for embodied model research and evaluation. The open-sourcing of MWA™-Phoenix gives the methods of motion-tactile learning and interaction prediction a chance to be reproduced and discussed. For Qianjue, practical use by frontier teams also provides important feedback for product R&D: which contact information is worth retaining, what kind of data is convenient for training, and how perception capabilities should adapt to specific tasks. Qianjue will continue to improve tactile products and the data foundation, using high-quality real tactile data to drive the continuous evolution of robot manipulation capabilities.