BOSTON, Oct. 24, 2026 (GLOBE NEWSWIRE) — DEEP Robotics has unveiled five research advances supporting the development of DEEPRobotics-Brain, its embodied intelligence architecture for real-world robot operation. Spanning physical prediction, navigation and manipulation, policy optimization, predictive memory, and data-driven evaluation, the research addresses a shared challenge: helping robots work reliably in environments that change.
The five research areas comprise the DR-WM-1.0 world model, the DR-VLA-2.0 embodied foundation model, the DR-APO policy optimization framework, the DR-Agent embodied agent, and the DR-Benchmark data and evaluation platform. Together, they explore how robots can anticipate action outcomes, translate instructions into motion, improve learned policies, revise outdated spatial beliefs, and measure progress using real-world data.
The five advances address interconnected challenges across this architecture. The research involves collaborations with researchers from the Massachusetts Institute of Technology, the University of California San Diego, the University of Illinois Urbana-Champaign, Tsinghua University, Zhejiang University, the Hong Kong University of Science and Technology, and the Beijing Institute of Technology. Together, the studies connect physical prediction, navigation and manipulation, policy learning, dynamic memory, and real-world data.
The research supports DEEP Robotics’ broader goal of developing robots that can respond to human instructions and practical operating conditions, with capabilities evaluated through simulation, real-robot experiments, and shared benchmarks.
DR-WM-1.0 Brings Physical Grounding to World Models
World models offer a way to evaluate actions before a robot executes them. Conventional robot training relies on repeated real-world trial and error, which is costly and time-consuming. Video-generative world models can simulate visually rich environments, but visual plausibility alone does not establish that a predicted action is physically feasible. A convincing video of a robot crossing snow, for example, may fail to capture slipping caused by low friction.
Developed with researchers from Zhejiang University, the Beijing Institute of Technology, Tsinghua University, MIT, and the University of Illinois Urbana-Champaign, DR-WM-1.0 addresses this gap by incorporating robot–terrain contact dynamics into action consequence prediction. The associated DR-WM study combines reconstructed scene geometry with dense physical priors for friction, roughness, and stiffness, calibrated using real-world action–response data.
The research draws on 2.58 million image–mask pairs and more than 100,000 real-robot clips. Its dual-ControlNet architecture conditions video generation on scene geometry and camera motion derived from simulated robot movement. During autoregressive rollouts, robot and controller states are carried forward, maintaining continuity with the underlying control loop.
Figure 1. DR-WM workflow and motion outcomes under different contact conditions.
This gives the robot a form of physical rehearsal: a way to estimate what might happen along a route and compare alternative actions. By connecting perception, physical prediction, and decision-making, DR-WM provides a testable approach to narrowing the gap between visual realism and physically grounded prediction.
An illustrative application is substation inspection after rain or snow. Before choosing between a slippery shortcut and a longer route, a robot could simulate both options, assess the likelihood of slipping, unstable motion, or limited progress, and weigh those risks against travel efficiency.
DR-VLA-2.0 Connects Human Experience with Robot Navigation and Manipulation
For robots to operate in streets, campuses, and delivery environments, navigation must extend beyond path planning. They need to follow instructions, avoid obstacles, anticipate changes, and move in ways that are compatible with human activity.
Collecting robot demonstrations through teleoperation is expensive and difficult to scale across diverse environments and interactions. Human egocentric web videos provide extensive examples of obstacle avoidance, yielding, and route selection. However, differences in camera viewpoint, imaging characteristics, and motion prevent direct transfer to robots, while the videos lack executable robot action labels.
Working with researchers from MIT, the University of Illinois Urbana-Champaign, Tsinghua University, and Zhejiang University, DEEP Robotics explores a learning pipeline that combines action-free pretraining on human video with adaptation using robot demonstrations. The associated DR-VLA study introduces EgoLynx, comprising 10,000 hours of curated egocentric web video and 2,000 teleoperated robot navigation clips.
Figure 2. DR-VLA overview. Human egocentric videos supervise understanding and prediction; robot demonstrations additionally supervise executable actions.
The central mechanism is a domain-invariant scene representation (DISR). It combines semantic, spatial, and relational information, allowing the model to focus on what is present, where it is, and how scene elements relate to one another. This structural representation provides an interface for transferring navigation knowledge across human and robot observations.
A Mixture-of-Transformers framework brings together vision-language understanding, latent world modeling, and a flow-matching action expert. Human video supplies supervision for scene understanding and prediction; robot demonstrations supply the action supervision needed to generate executable trajectories. The approach is intended to support socially compliant behaviors such as waiting, bypassing obstacles, and yielding to pedestrians.
On the held-out EgoLynx-Robot test split, DR-VLA improves the Quadruped Navigation Score (QNS) by approximately 5% and reduces open-loop final heading error by 14.6%, each relative to the best competing result for that metric. QNS combines safety, instruction compliance, comfort, and progress. These results point toward more natural robot navigation in shared spaces, including last-mile delivery in residential communities and campuses.
Manipulation presents a related challenge: translating language into reliably controllable motion. In existing vision-language-action (VLA) models, strong visual priors can overshadow language instructions. Changing a command may fail to redirect an arm, while contact-rich tasks such as insertion and wiping require motion to adapt to force feedback.
DEEP Robotics and its research collaborators address this problem with Atomic Motion Coordinate (AMC), a framework for language-steerable and force-responsive manipulation. AMC defines 13 interpretable motion atoms for each arm: positive and negative translations along three axes, positive and negative rotations about three axes, and a hold action. An initial training stage excludes visual inputs to align language with geometric motion. Subsequent stages use visual context to realize those motion tendencies and incorporate force feedback to adjust generated actions within the same coordinate framework.
Figure 3. With images and robot states fixed, opposite motion-atom prompts produce opposing end-effector translation and rotation responses.
AMC achieves opposite-atom separation rates of 92.5% for single-atom prompts and 83.1% for two-atom prompts. In real bimanual tasks with unseen layouts, normalized task progress reaches 87.8%. Adding force feedback raises task progress to 78.5% for the drawer-and-plug task and 75.2% for vase wiping. The evaluations demonstrate improvements in language steerability, generalization to held-out configurations, and contact adaptation relative to the tested baselines, including fine-tuned π0.5 and a reproduced LA4VLA-style approach.
Potential applications include electrical cabinet operation, connector insertion, and component alignment. The results support more interpretable control of manipulation.
DR-APO Improves Reinforcement Learning Across Policy Representations
As navigation and manipulation models grow more capable, training methods must support increasingly diverse action representations. Existing reinforcement learning approaches are often tailored to a particular policy parameterization, limiting their reuse across models and tasks.
Developed by DEEP Robotics and Zhejiang University, DR-APO (Decoupled-Representation Advantage-Guided Policy Optimization) separates policy improvement from the underlying policy representation. The framework provides a shared optimization principle for Gaussian, diffusion, and flow-matching policies, supporting both continuous-control learning from scratch and post-training of pretrained embodied navigation policies.
Figure 4. DR-APO combines advantage-guided target improvement with representation-compatible action fitting for continuous control and navigation post-training.
Its advantage-guided target improvement mechanism constructs improved action targets using local advantage gradients at executed actions and advantage-based reweighting. Policy improvement is formulated as target optimization independent of the policy parameterization. This avoids explicit policy-likelihood evaluation for actor updates and does not require critic-based selection among multiple candidate actions at inference time.
Representation-compatible direct action fitting then transfers these targets to Gaussian, diffusion, and flow policies. This is designed to prevent local improvement signals from being diluted during multistep generation.
Across five MuJoCo-v5 continuous-control tasks, DR-APO achieves the best result in nine of ten diffusion-policy comparisons and all ten flow-policy comparisons reported in the study. It delivers the best aggregate policy-only performance for both policy families, without inference-time multi-candidate critic selection, while remaining competitive with Soft Actor-Critic (SAC) for Gaussian policies.
In embodied visual navigation post-training, DR-APO raises pretrained NavDP’s simulated success rate from 70.33% to 89.55%, an increase of 19.22 percentage points. In real-world navigation tests on the DEEP Robotics LYNX M20 wheel-legged robot, success rises from 25% to 65%, a gain of 40 percentage points. These findings demonstrate the framework across continuous-control learning and navigation post-training, providing a basis for further research on other embodied policies.
DR-Agent Updates Spatial Beliefs as the World Changes
Reliable long-term operation requires robots to reason about changes that occur outside their field of view. Objects are moved in homes, hotel spaces are rearranged, and inventory shifts in warehouses. A previously accurate memory can become obsolete before the robot reaches its destination.
Developed with researchers from the University of California San Diego and Zhejiang University, DR-Agent draws on EvolvingNav, a navigation method built around predictive 4D belief. The method addresses persistent navigation when targets may move while unobserved, including during the robot’s journey.
EvolvingNav constructs a time-indexed belief from timestamped 3D object histories. An event-driven predict–observe–replan filter estimates where a target is likely to be at the expected inspection time, incorporates new visual evidence, and revises the robot’s plan. This enables memory and location beliefs to be updated as the environment changes.
Figure 5. Overview of the predictive belief and navigation framework underlying DR-Agent.
The accompanying EvoWorld-Bench benchmark is grounded in human activity traces and comprises 54 simulated scenes, 803,680 task instances, and approximately 239,000 dynamic state changes. Experiments show improvements in navigation success and search efficiency over the evaluated baselines, with particularly clear gains when temporal patterns can be learned. Evaluations on FindingDory and GOAT-Bench further demonstrate applicability to embodied navigation tasks.
Potential uses range from sustained autonomous service in homes and shared spaces to object search in warehouses and inspection-route adjustment in industrial facilities. Predictive, revisable memory helps address a central deployment requirement: maintaining useful behavior over repeated tasks in changing environments.
DR-Benchmark Unifies Evaluation on Real-World Robot Data
All of these capabilities depend on data that connects perception, intent, and continuous robot motion. Existing robot and language-navigation datasets often provide only part of this information. Perception, navigation, future prediction, and reconstruction are also frequently evaluated using different datasets and protocols, making consistent comparison difficult.
DEEP Robotics developed DR-Benchmark with researchers from the University of Illinois Urbana-Champaign, the Hong Kong University of Science and Technology, Zhejiang University, Tsinghua University, and MIT. At its core is LynxScenes, a large-scale real-world dataset and benchmark for outdoor wheel-legged robot operation.
Figure 6. Lynx Scenes overview, including real-robot data collection, multimodal observations, spatial and behavioral annotations, and four downstream tasks.
Collected using the DEEP Robotics LYNX M20, LynxScenes spans 833.6 hours and 3,294.7 kilometers of real-world operation. It contains approximately 100,000 navigation clips and 3 million annotated frames, with supervision covering 3D objects, road structures, navigation instructions, and expert demonstration trajectories. The annotations encompass 27 object categories, four road-structure categories, and 20 navigation instructions organized into three levels.
The benchmark brings four complementary tasks onto a shared real-world dataset: instruction-conditioned trajectory prediction, bird’s-eye-view (BEV) perception of 3D objects and road structures, future scene prediction, and multi-view scene reconstruction. The research evaluates 25 representative models, enabling more consistent analysis across navigation, perception, world modeling, and reconstruction.
Applications include studying when delivery robots should proceed, wait, or detour in shared spaces; analyzing navigation through narrow passages and temporary obstacles during industrial inspection; and replaying recorded scenarios to diagnose model behavior. DR-Benchmark provides a basis for model development and comparison, while linking real-world operating data with subsequent research and improvement.
Together, these research advances support the development of DEEPRobotics-Brain and its underlying architecture. They address complementary requirements for real-world robotics: physically grounded prediction, instruction-driven navigation and manipulation, policy improvement, revisable spatial memory, and consistent evaluation on recorded robot experience.
DEEP Robotics is pursuing this coordinated research program to support development across robot embodiments. By connecting these research directions with verifiable methods and shared benchmarks, the company aims to advance embodied AI toward reliable industrial applications and sustained human–robot collaboration.
The research preprints are available in the research section of the DEEP Robotics Developer Center: https://developer.deeprobotics.cn/research/ .
Media Contact
Company: DEEP Robotics
Contact: Vivian Chen
Email: chenlingjia@deeprobotics.cn
Website: https://www.deeprobotics.cn/en
Disclaimer: This content is provided by DEEP Robotics. The statements, views, and opinions expressed in this content are solely those of the content provider and do not necessarily reflect the views of this media platform or its publisher. We do not endorse, verify, or guarantee the accuracy, completeness, or reliability of any information presented. This content is for informational purposes only and should not be considered financial, investment, or business advice. All investments carry inherent risks, including the potential loss of capital. Readers are strongly encouraged to conduct their own due diligence and consult with a qualified financial advisor before making any investment decisions. Neither the media platform nor the publisher shall be held responsible for any inaccuracies, misrepresentations, or financial losses resulting from the use or reliance on the information in this press release. Speculate only with funds you can afford to lose. In the event of any legal claims or concerns regarding this article, we accept no liability or responsibility. Globenewswire does not endorse any content on this page.
Legal Disclaimer: This media platform provides the content of this article on an “as-is” basis, without warranties or representations of any kind, express or implied. We assume no responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained herein. Any complaints, copyright issues, or inquiries regarding this article should be directed to the content provider listed above.
Photos accompanying this announcement are available at
https://www.globenewswire.com/NewsRoom/AttachmentNg/d9904360-3a8b-4d37-9a4c-617f11d853e2
https://www.globenewswire.com/NewsRoom/AttachmentNg/1c879e9e-e245-4fa4-8edc-bd16f32aa8fe
https://www.globenewswire.com/NewsRoom/AttachmentNg/58acf734-f667-40ed-b8dc-79ae827b4276
https://www.globenewswire.com/NewsRoom/AttachmentNg/a633fd25-bbf8-427e-b0a4-b88bd249222a
https://www.globenewswire.com/NewsRoom/AttachmentNg/a2325259-2967-45dc-a9c5-91eccb5ddd21
https://www.globenewswire.com/NewsRoom/AttachmentNg/c4e24e92-21af-4333-83a6-45e6d4a0a231
