DEEP Robotics sets out five studies for outdoor robots

The release's DR-WM-1.0 is the paper titled DR-WM[1][2]. DR-VLA-2.0 is the paper titled DR-VLA, and the same section of the release also describes Atomic Motion Coordinate[1][3][4]. DR-Agent is the system the EvolvingNav paper proposes[1][6]. DR-Benchmark is LynxScenes[1][7]. DR-APO keeps its name in both places[1][5].

The release lists collaborators at 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[1]. Each preprint carries its own author list.

DR-WM predicts how a quadruped's action will turn out on outdoor ground, then generates the view that motion would produce[2]. The paper says Physcene-Image contains 2.58 million unique RGB-mask pairs: 1.5 million on-robot, 690,000 indoor, 44,600 off-road and 350,000 from the web. Physcene-Robot is more than 100,000 robot clips with images, robot state and control signals[2]. The release's counts, 2.58 million image-mask pairs and more than 100,000 real-robot clips, match those totals[1].

The friction, roughness and stiffness channels are not a sensor reading on every pixel. The paper says masks from 32 semantic classes are filled from a class-to-property lookup built on language-model priors. A calibration learned only on the training set then maps those channels to simulator parameters, using real robot action-response data, and the mapping stays fixed at test time[2]. The release says the priors are calibrated with real-world action-response data. It does not describe the lookup[1].

Video generation uses a dual-ControlNet, conditioned on scene geometry and on camera motion taken from the simulated robot[1][2]. Across 90 scenarios the paper calls challenging, an 8B vision-language model that can read the rollouts beats an unassisted 32B vision-language model on five of six metrics[2]. The release's example, a robot choosing a route through a substation after rain or snow, is an illustration of that idea. It is not identified as one of the 90 scenarios[1].

DR-VLA trains on EgoLynx: 10,000 hours of curated egocentric web video and 2,000 teleoperated robot navigation clips from real deliveries[3]. The release prints the same two figures[1]. The paper says the understanding and world-model experts are pretrained on the web video without action labels, then adapted to robot images, with a flow-matching action expert trained on the robot clips[3].

On the held-out EgoLynx-Robot split, DR-VLA improves its Quadruped Navigation Score by about 5% and cuts open-loop final heading error by 14.6%, each against the best other result on that metric[3]. The release repeats both figures and says the score combines safety, instruction compliance, comfort and progress[1]. Trained separately on NAVSIM v1, without the web pretraining, the paper says DR-VLA scores 92.1 PDMS[3]. The release does not mention NAVSIM.

Atomic Motion Coordinate is the manipulation result inside that same section of the release[1]. The paper gives each arm 13 motion atoms: positive and negative translation on three axes, positive and negative rotation on three axes, and a hold[4]. Across 7,520 offline horizon interventions, opposite-atom separation is 92.5% for one atom and 83.1% for two, against 39.1% and 24.0% for a reproduced LA4VLA-style method[4]. The release prints 92.5% and 83.1% and does not print the comparison rates[1].

On the robot, the paper reports 50 trials per task. Force-free AMC reaches 87.8% normalized progress on out-of-distribution fruit tasks, up from 60.5% for a fine-tuned π0.5[4]. Force history then moves the drawer-and-plug task from 59.0% to 78.5% and vase wiping from 71.6% to 75.2%. Those from-scores are force-free AMC, not π0.5, which the same table lists at 55.0% and 65.2% on the two contact tasks[4]. The release says unseen-layout progress reaches 87.8%, and that force feedback raises the drawer-and-plug task to 78.5% and vase wiping to 75.2%[1]. It does not give the trial count or the from-scores.

DR-APO is a reinforcement-learning update meant to improve Gaussian, diffusion and flow-matching policies through the same target step, without scoring several candidate actions with a critic at inference time[5]. On five MuJoCo-v5 tasks, the paper says it leads in 9 of 10 diffusion comparisons and all 10 flow comparisons, and that it stays competitive with Soft Actor-Critic on Gaussian policies[5]. The release states the same 9-of-10 and 10-of-10 counts[1].

For navigation, the paper and the release both say post-training raises NavDP's simulated success from 70.33% to 89.55%, and real-world success from 25% to 65%[5][1]. The simulation table is a wheeled robot in InternScenes. There, DR-APO's success-weighted path length is 74.55, below X-NavDP at 77.67, while its success rate is 89.55 against X-NavDP's 88.63[5]. The release does not mention path length.

The real-world table is 20 scenarios: 25% for pretrained NavDP and 65% after DR-APO[5]. The paper calls that robot a DEEP Robotics M20 quadruped[5]. The release calls it the LYNX M20 wheel-legged robot[1]. The table does not say how many runs sit behind each percentage.

EvolvingNav, which the release calls DR-Agent, keeps a time-indexed belief about where a target is likely to be when the robot arrives, including targets that move while unobserved[6][1]. The paper's EvoWorld-Bench spans 54 scenes and 4,860 scene-days, with approximately 239,190 dynamic state changes and 803,680 task instances[6]. The release says 54 simulated scenes, 803,680 task instances and approximately 239,000 dynamic state changes[1]. The paper lists a code repository, a Hugging Face demo and a project page[6].

LynxScenes is the dataset under DR-Benchmark[1][7]. Collected on a LYNX M20 the paper calls wheel-legged, it contains about 100,000 non-overlapping 30-second clips, 833.6 hours, 3,294.7 km and 3.0 million annotated frames, with four fisheye cameras, two LiDARs, an IMU, odometry and robot state[7]. Annotations include 20.82 million oriented 3D boxes across 27 categories, 13.89 million road-structure annotations, four road-structure classes, and 20 navigation instructions in three layers[7]. The release's scale figures match the hours, the distance, the approximate clip count and the 3 million frames[1].

The paper evaluates 10 instruction-following policies, 4 bird's-eye-view perception models, 6 world models and 5 reconstruction methods on the same sequences[7]. That is the release's count of 25 models[1]. LynxScenes also defines a Quadruped Navigation Score, with components that include collision avoidance, walkable-area compliance, clearance, progress and instruction adherence[7]. DR-VLA's score of the same name is the one the release defines as safety, instruction compliance, comfort and progress[3][1]. The two papers do not say the scores are the same number.

The LynxScenes paper says the dataset, the splits and the evaluation toolkit are planned for public release[7]. The 24 October release says the preprints are on the developer center. It does not say the recordings can be downloaded[1].

The news on 24 October is a name. DEEPRobotics-Brain is the label the company puts on outdoor-robot research that was already on arXiv, from a world model to a logged benchmark[1]. The scale that is easiest to check is LynxScenes: 833.6 hours and 3,294.7 km on a LYNX M20, with the files still described as forthcoming[7]. The real-robot change from 25% to 65% is one table of 20 scenarios, on a machine the paper calls an M20 quadruped and the release calls a LYNX M20 wheel-legged robot[5][1]. DR-WM's terrain channels start as a language-model lookup over semantic classes, then a calibration fit to robot motion[2]. The figures above are the authors' own evaluations.