“Go to the chair, find the door, then return to the cone you passed earlier.” The last clause is harder than it sounds. The cone may no longer be visible. A robot has to remember which cone was meant, relate an old camera view to its current heading, and keep moving consistently while its body fails to execute every small command exactly. Compass 1 is built around this gap between recognizing a scene and navigating through a history of scenes.

Led by Archon Robotics with HKU, Shanghai Innovation Institute, Spatial Robotics, and Huawei, the work combines sparse visual memory with a video-action foundation model. Its most useful contribution is the training and execution recipe that makes history actionable: structured questions supervise stage, exploration, and target retrieval, while a spatial canvas records where keyframes were observed. The camera and reconstruction pipeline explain how that recipe gets its supervision. [1] [2]

Evidence checked on 14 November 2026, based on Compass 1 paper v1, its official research page, and the release repository. Numbers are author-reported, not independently reproduced. This is Archon’s Compass 1 navigation model, distinct from other projects named COMPASS. We distinguish offline geometry labeling, online model inputs, and the physical control system throughout.

1. Which camera does Compass 1 actually use?

Stage / platformReported camera and heightWhat this establishes
Self-collected training videoDJI Action 4, stabilization enabled; mainly around 1.0 mHuman-captured navigation video; not 500 h of robot teleoperation logs
Unitree Go2 main evaluationDJI Action 4 bonded to the head, about 0.5 mMain 320-trial benchmark tests a lower, moving viewpoint
AgileX wheeled / Booster K1 humanoidDJI Action 4, 0.35 / 0.85 mCross-embodiment tests retain the same ground-camera family
Spatial Robotics droneNative industrial camera, about 1.5 mExact camera model is not identified in the paper

Source: Sections 3 and 4.1. The policy’s visual input is monocular RGB; the experiment is not RGB-D navigation. However, every evaluated physical method relies on odometry for closed-loop velocity execution, and the main model runs remotely on an NVIDIA RTX PRO 6000 GPU. “Single camera” describes the visual modality, not a robot built from a camera and a network with no state estimation, controller, or communications dependency. [2]

The paper does not specify a complete capture profile covering resolution, lens mode, field of view, exposure, stabilization settings, and calibration files. Its 4 FPS example explains annotation-time temporal sampling; it is not a claim about the Action 4’s recording frame rate. For reproduction, those settings matter because stabilization and cropping can change the effective imaging geometry. This is an engineering concern to test, not a reported source of failure in the study.

2. Its 3D reconstruction primarily produces action labels

Compass 1 does not describe a WorldEngine-style 3D Gaussian scene that is rebuilt and rendered for every navigation episode. Its data pipeline combines π³ (Pi3) with MoGe-2 to recover metric motion from monocular video. Pi3 estimates camera poses and local point maps across views; MoGe-2 supplies a learned metric-scale estimate. Camera motion can then become continuous body-frame navigation increments: forward displacement Δx, lateral displacement Δy, and yaw change Δyaw. [1] [2] [4] [5]

Geometry componentRole in the pipelineBoundary of the evidence
Pi3Multi-view camera poses and point maps without a fixed reference viewReconstructed motion is estimated supervision, not measured robot ground truth
MoGe-2Recover metric-scale geometry from monocular imagesA learned scale prior can be wrong; scale-error statistics are not reported here
Continuous action extractionConvert trajectory changes into [Δx, Δy, Δyaw] labelsPaper does not publish a full scale-fitting / calibration implementation yet
History Representation MapDraw keyframe positions, IDs, headings, and past trajectory on a 2D canvasNot a dense 3D occupancy map or collision certificate

This distinction is central. Offline geometry provides labels and spatial structure for learning. The online History Representation Map (HRM) is an image of the traveled history, not the original 3D point cloud. A red current-pose marker, heading arrows, and indexed past views make return targets interpretable to the existing visual backbone. The public description does not establish a deployment-time dense reconstruction service, loop closure, or a full obstacle map. [2]

The authors filter trajectories with implausibly large action scale, excessive roll/pitch, or simultaneous spikes in translation and yaw near tight turns. This is sensible quality control, but filtered imitation data can omit precisely the awkward motions that robots encounter during recovery. The paper does not report rejection yield, metric trajectory error, annotation person-hours, or a matched cost against conventional capture. A semi-automated pipeline is credible; a measured cost advantage still needs those numbers. [2]

3. The data contribution is alignment across geometry, language, and stages

From monocular video to navigation supervision

Recover and select

Action 4 / web video → Pi3 geometry + MoGe-2 scale
Extract continuous motion → remove noisy trajectories
Keep turns densely and straight motion sparsely

Align instruction and targets

Motion primitives + video → stage-aligned language
Keyframes → target stages → refine temporal boundaries
Anchor final target instance → ground backward

Verify and supervise

Review self-collected boundary clips and action plots
Correct language → rerun affected annotation stages
Train QA + video + action jointly

Paper Sections 3.1–3.6; human verification is applied only to self-collected trajectories

The corpus contains over 500 hours of self-collected real-world video and over 2,000 hours of annotated YouTube video. Training samples the two sources at 85% and 15%, respectively. The larger web inventory therefore does not dominate optimization by raw duration. Human checks apply only to the self-collected portion. These are video-derived action labels rather than uniformly instrumented robot demonstrations with measured forces, collisions, and actuator responses. [2]

The annotation design is more specific than asking a language model to caption a walk. Motion primitives constrain turning and stage boundaries; Doubao-Seed-Pro 2.1 creates instructions from video plus that action timeline. Qwen-3.8-27B assigns target stages and refines boundaries. The final view of each stage anchors the intended target instance, whose grounding is propagated backward. This reduces confusion between similar chairs or bins. The quality of these derived labels remains a dependency, especially in web video that lacks manual verification. [2]

4. What the model changes: memory needs a training objective

Compass 1 uses Cosmos3-Edge-4B as its foundation backbone. Sparse historical keyframes, the HRM, the current observation, instruction stages, and the previous intention enter shared multimodal attention. The autoregressive language route answers structured questions; the diffusion route jointly predicts twelve future video frames and twelve continuous navigation actions. The 4B label describes the foundation model, not a separately audited parameter total for every system component. [1] [2]

Supervision categoryQuestions / state being learnedExecution problem it targets
Task-stage groundingCurrent stage, stop decision, forward/left/right intentionPremature stop and unfinished multi-stage instructions
Exploration consistencyPreviously explored directions + intention propagated to the next stepLeft-right oscillation and inconsistent searching
Spatiotemporal target retrievalWhich historical keyframe shows the target, and where that view was relative to nowRemembering an object without knowing how to return

The mechanism addresses a practical failure of adding memory to a large model: the model can ignore it. HRM supplies a readable spatial reference, and the questions force the training signal to use it. Action-guided sparsification also reduces a train–deployment mismatch. When a robot repeatedly issues tiny motions that barely move its body, many consecutive frames are nearly identical. Selecting online keyframes by traveled distance and rotation better matches the training representation than appending every camera frame. [2]

In annotation, straight segments are sampled at three-second intervals and turning regions more densely when accumulated yaw exceeds 30 degrees, compressing a 240-frame, one-minute example to roughly forty keyframes. The paper’s nearly two-minute retrieval evidence is meaningful, but it is a bounded history window. It does not demonstrate a persistent building-scale map, multi-day memory, or automatic correction of long-term odometry drift. [2]

Shared model, asynchronous execution

Inputs and shared context

Current RGB + indexed history + HRM + instruction
Previous intention becomes context for the next decision
Shared Cosmos3-Edge multimodal attention

Video-action route

Joint diffusion predicts 12 frames + 12 [Δx, Δy, Δyaw] increments
Begin trajectory execution without waiting for current QA answers
Odometry-backed controller realizes motion

Language route

Autoregressively predict stage, intention, stop, and retrieval answers
Intention informs the following decision
Stop signal asynchronously interrupts the controller

The video-action route is not a serial planner consuming all current QA answers; the official page explicitly separates the routes

The model is trained end-to-end for 60,000 steps on 32 H200 GPUs with global batch size 256; the VQA, video, and action losses have weight ratio 3:1:1. That is supervised co-training, not reported online navigation RL. No total wall time or complete GPU-hour bill is provided. The video-removal ablation tests whether that training route helps; it does not establish that the robot searches a tree of imagined worlds or that generated images certify collision-free actions. [2]

5. How much does the real-robot result establish?

Main Go2 testCompass 1Comparison / interpretation
Overall: 8 task types × 4 scenarios × 10 trials254/320 = 79.4%AwareVLN 17.8%; same authors’ benchmark
Initially visible target98%Rounded project-page task rate
Multi-stage instruction50%All four tested baselines 0%; substantial remaining failure
Basic history return98%Return target can become visible after turning
Persistent history return58%All four tested baselines 0%; return target initially occluded

Source: official results and Sections 4.1–4.2. Each task type pools forty trials across Office, Complex, Terrace, and Street; displayed task percentages are rounded. The overall improvement over the strongest reported baseline is 61.6 percentage points, not 61.6% relative. All baselines are adapted to the authors’ data, and results select the better of original and fine-tuned checkpoints. This is stronger than comparing against unadapted weights, but architectures, action encodings, and reasoning supervision still differ. For example, AwareVLN’s adaptation receives action-only supervision rather than newly annotated reasoning. It is a system comparison, not an isolated memory experiment. [1] [2]

The four-embodiment claim needs its own denominator. Additional history-backtracking tests report 9/10 on the wheeled robot, 7/10 on the humanoid, and 7/10 on the drone, without fine-tuning. These thirty trials are separate from the Go2 main benchmark; they do not establish 79.4% across all four platforms. Transfer spans camera height and body motion, while retaining an abstract planar displacement/yaw interface and platform-specific control. It is not a learned universal motor policy or demonstrated unrestricted 3D flight. [1] [2]

6. The ablations show a useful combination, not one magic module

Variant, Office + StreetReal success rateWhat changes
Full Compass 178.8%Reference on the same 160-trial subset
Without HRM53.1%Remove spatial canvas; retain structured QA
Without spatiotemporal target retrieval56.3%Remove explicit historical target-use supervision
Without task-stage grounding55.0%Replace learned stopping / stage supervision as described in paper
Without exploration consistency61.3%Remove direction-history QA and intention propagation
Without video generation46.9%Use a direct diffusion action head without that auxiliary route
Without YouTube data65.0%Train on self-collected data only

Each variant has 160 real trials, twenty per task across two scenarios. Its reference is 78.8%, not the full four-scenario 79.4%. The declines support the usefulness of explicit memory and its supervision, and of the joint predictive training route under this recipe. They cannot be added into independent percentage-point contributions. Several variants alter interfaces or supervision together, and the report does not provide repeated-training uncertainty for these comparisons. [1] [2]

Simulation provides a different scale of evidence. On R2R-CE Val-Unseen, Compass 1 reports SR 70.1 and SPL 63.2, versus LightNav-0’s 68.5 and 62.8 with monocular RGB. The physical gap is much larger than that simulation gap, which makes execution consistency a plausible explanation worth investigating. On IndoorUAV, all methods are fine-tuned, upward/downward-action trajectories are excluded, and Compass 1 SR is 18.09. Those conditions rule out reading the drone result as solved zero-shot aerial navigation. [1] [2]

7. “17–28 Hz” is action throughput, not a new visual decision every 36 ms

The authors report producing twelve continuous actions in 0.43 seconds initially and 0.71 seconds at 63 keyframes on PRO 6000. Dividing twelve by those times gives the quoted effective 28/17 Hz. The corresponding batch-ready decision rate is approximately 2.3/1.4 batches per second, calculated from the reported times. These are compute measurements, not end-to-end camera-to-motor latency, and they do not include network delay merely because execution is asynchronous. [2]

The diffusion route can start motion before the language head finishes. The previous intention informs the next decision, and a stop answer interrupts the controller asynchronously. KV caching and CUDA Graphs reduce repeated work. This is a practical scheduling contribution. Deployment still needs to measure stale-command age, lost packets, interrupt delay, braking distance, and behavior when the cloud is unavailable. The published results do not establish onboard edge-GPU parity or a verified local emergency-obstacle layer. [2]

8. How much of end-to-end robot navigation is solved?

Our assessment is a strong controlled real-world demonstration of history-aware navigation, with a substantial advance on the authors’ long-horizon task suite. The unique contribution is the alignment of representation, supervision, and execution: indexed visual memory plus a spatial canvas, questions that make the model use that memory, and an action interface that tolerates imperfect body execution. Pi3, MoGe-2, and Cosmos3-Edge are reused foundations; the contribution is the navigation system built around them.

It has not solved persistent autonomy. The main benchmark still has 66 failures in 320 trials; multi-stage execution succeeds about half the time and persistent backtracking about 58%. Reported failures include dynamic occlusion, missed targets during rotation, reflection and lighting confusion, occasional collision, and losing the return heading around obstacles. Four selected scenarios do not measure multi-floor routing, crowded-site safety over a shift, long-term mapping drift, or mission recovery with bounded intervention. [2]

Reproducibility is also pending. The abstract describes an open-source dataset, but the official repository schedules dataset, deployment code, and real-world checkpoints for December 2026. As checked, it contains the research introduction and assets rather than the complete released system. The public Apache-2.0 repository license should not be assumed to license unreleased videos or weights, and web-video rights need their own review. Training duration, collection cost, and continuous operating cost remain insufficiently reported. [3]

For a robot navigation team, follow this work now and test its memory ideas before replacing the whole stack. Hold perception and control fixed, then compare ordinary frame history, indexed keyframes, and HRM plus retrieval supervision on return-to-landmark tasks. Evaluate on unseen sites and camera settings, count complete missions, collisions, interventions, and time, and stress the communication path. The next decisive advances would be reliable geometry under drift, uncertainty-aware recovery after target loss, local deployment with measured total latency, and independent replication after release. Compass 1 makes remembered experience more useful for navigation; a safe navigation product still has to manage the failures around that capability.

What the reconstruction comparison can establish

The geometry survey explains why Pi3X or MapAnything is a candidate label generator, not an established replacement for Pi3 + MoGe-2. The roadmap therefore starts with reference-measured motion error and a matched downstream policy. Neither a different geometry backbone nor removing a dense online map proves that memory is unnecessary. Archon’s Compass 1 also differs from NVIDIA’s COMPASS agent framework; their results must remain separate.

Sources and implementation status

  1. Compass 1 official research page — method, task counts, ablations
  2. Compass 1 paper v1 — hardware, labeling, model, and evaluation
  3. Official Compass repository — December release schedule
  4. Pi3 official code — camera geometry and point maps
  5. MoGe-2 official code — metric monocular geometry
  6. Original Archon Compass 1 research URL