Compass 1: Towards Open-World and Persistent Embodied Navigation from Monocular RGB Videos

Hai Zhang, Siqi Liang, Yanjun Guo, Shoufeng Zhang, Yunhao Wang, Hangyu Du, Haoguang Mai, Hongchen Li, Jiazhi Yang, Jialong Zeng, Yu Hu, Chen Chen, Gangcheng Jiang, Zishun Zhuo, Lirui Zhao, Zhigang Li, Yu Qiao, Zhenjie Yang, Fu Zhang, Hongyang Li

November 2026

Zero-shot run on the humanoid in an indoor corridor.

TL;DR A language-conditioned navigation model that sees only a monocular RGB camera and remembers almost two minutes of its route, using sparse keyframes plus a drawn map of its own past trajectory. In 320 zero-shot real-world trials on a quadruped it succeeds 79.4% of the time, against 17.8% for the best baseline, and it also runs on a wheeled robot, a humanoid and a drone.

What they do

Navigation models trained in simulation often fail on real robots: they stop early, turn the wrong way, get stuck in left-right deadlocks and barely use their history. Compass 1 trains on real monocular RGB video and adds a long memory, sparse keyframes plus a drawn map of its past path, without changing the backbone architecture.

How it works

Backbone. Cosmos3-Edge-4B with an autoregressive language head for reasoning and a diffusion head that generates future video and 12 actions (dx, dy, dyaw).
History. Keyframes are taken at fixed distance and angle thresholds. The History Representation Map draws them, with indices and heading arrows, as an extra input image.
VQA supervision. Three question types: which task stage is running and whether to stop; which directions are already explored; and which keyframe shows the target and in what direction it lies. Without these questions the model ignores the map.
Data. Over 500 h of self-collected camera video plus over 2,000 h of YouTube video, labeled semi-automatically with actions from π3 and MoGe-2.

Results

On a Unitree Go2 with its camera at 0.5 m (training is mostly at 1.0 m), 8 tasks × 4 scenes × 10 trials: 79.4% overall success. Four baselines, each scored at the better of its released and its fine-tuned weights, reach 17.8%, 11.6%, 6.3% and 0.3%.
History backtracking: 78% vs 4%, with 98% on basic and 58% on persistent backtracking. Exploration: 84%. Instruction following: 78%, but only 50% on multi-stage routes.
In crowded streets the best baseline drops by nearly 65%; Compass 1 drops only slightly.
Inference runs at about 28 Hz at the start and 17 Hz at 63 keyframes.

Limitations

Multi-stage routes fail after turning errors accumulate or when targets blend into glass reflections or harsh lighting. Persistent backtracking can lose its heading when obstacles block the way back. Camera jitter lowers success on the humanoid and the drone. Inference runs on a remote GPU, and velocity control relies on odometry.

Videos

The humanoid walking down a corridor toward a yellow chair.

Figures

From Figure 1: Open-world evaluation tasks: instruction-following (IFN), beyond-the-view (BVN) and history-backtracki... — From Figure 1: Open-world evaluation tasks: instruction-following (IFN), beyond-the-view (BVN) and history-backtracking (HBN) navigation. Outer panels show the reference paths; inner panels overlay zero-shot real-world runs on a wheeled robot, quadruped, humanoid and drone with camera heights of 0.35, 0.5, 0.85 and 1.5 m.

From Figure 2: (a) The history representation map draws past keyframes and the past trajectory on a 2D image, with me... — From Figure 2: (a) The history representation map draws past keyframes and the past trajectory on a 2D image, with merged indices and heading arrows. (b) A unified understanding-and-generation backbone takes the instruction, the map and the keyframes. Structured VQA supervision drives history-aware reasoning, and a diffusion video-action head generates action trajectories.

From Figure 5: Open-loop history backtracking over a nearly two-minute trajectory. The model finds the keyframe index... — From Figure 5: Open-loop history backtracking over a nearly two-minute trajectory. The model finds the keyframe index showing the target and its direction from the current position, and updates both when the target is seen again.

From Figure 7: Zero-shot real-world success rates: overall, per task, per task category and per scenario against base... — From Figure 7: Zero-shot real-world success rates: overall, per task, per task category and per scenario against baselines, plus cross-embodiment transfer and ablations.

Abstract

Embodied navigation requires agents to achieve persistent spatial awareness and long-horizon execution stability in open-world environments. However, existing navigation architectures struggle to sustain both capabilities, remaining effective only over a handful of recent frames. We present Compass 1, the first embodied navigation foundation model to unify language understanding and video generation while enabling memory backtracking through nearly 2mins of visual history. Without altering the underlying architecture, we introduce History Representation, which integrates sparse visual keyframes with an explicit History Representation Map to preserve long-term spatiotemporal context. To effectively leverage this history, we further introduce History Utilization, a structured VQA supervision protocol comprising three categories: task-stage grounding, exploration consistency, and spatiotemporal target retrieval, enabling map-grounded reasoning and stable persistent long-horizon navigation. To fuel this system, we establish a complete annotation pipeline for monocular RGB videos and open-source a massive dataset featuring over 500 hours of high-quality real-world interactions and over 2,000 hours of annotated web videos. Trained mainly on the distribution of 1m height camera, Compass 1 generalizes seamlessly across diverse embodiments and scenarios, ranging from indoor environments to pedestrian-dense streets, spanning wheeled robot at 0.35m, quadruped dog at 0.5m, humanoid at 0.85m, and aerial drone at 1.5m. Compass 1 achieves a dominant overall success rate of 79.4% across open-world evaluations, compared to 17.8% for the strongest baseline. It demonstrates exceptional zero-shot execution exceeding 90% on foundational tasks, a 250% boost in exploration, and an unparalleled leap from 0% to 58% in challenging persistent history backtracking.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong