Pith. sign in

REVIEW 4 major objections 6 minor 23 references

Milo is claimed to be the first open-source, roughly $2,000 quadruped robot guide dog that performs fully autonomous, map-free indoor and outdoor collaborative navigation while measuring the handler's position through a sensorized handle.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Milo is an open-source, fully onboard robotic guide dog that navigates unseen indoor/outdoor paths while explicitly modeling the handler's position, with preliminary real-world tests against a handler-unaware costmap baseline.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A genuine open-source platform contribution with honest limitations, but the empirical support for the headline claims is thinner than the text suggests. the 4 major comments →

arxiv 2607.19530 v1 pith:BE35ULCO submitted 2026-07-21 cs.RO

Milo, a Fully Autonomous Indoor/Outdoor Robotic Guide Dog

Florian Golemo , Joanna Wolski , Joel Ruben Antony Moniz , Christopher Pal This is my paper
classification cs.RO
keywords robotic guide dogassistive roboticsquadruped robotbird's-eye-view navigationreinforcement learningobstacle avoidanceblind and low-vision mobilityzero-shot sim-to-real
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that a guide-dog-style navigation robot does not need an expensive animal, a pre-scanned map, or an external computer: by fusing onboard camera, lidar, and a handle-mounted encoder into a bird's-eye-view image, a small robot can follow walkable paths and avoid obstacles and pedestrians in places it has never seen. The authors claim Milo is the first open-source platform at about $2,000 that can fill the basic collaborative navigation role of a guide dog, meaning it takes turn and speed commands from the handler while shielding the handler from collisions. In real indoor and outdoor courses, they compare the learned policy against a costmap baseline and report smoother path-following and fewer handler collisions, at the cost of some robot collisions with the environment. A sympathetic reader would care because it is a concrete, reproducible, low-cost testbed for an assistive capability that has been blocked mostly by cost and by robot autonomy.

Core claim

The central claim is that a bird's-eye-view (BEV) semantic map can serve as the shared language between a fast 2D simulator and a real robot: the policy is trained entirely in a GPU-accelerated BEV simulator that renders roads, walls, obstacles, pedestrians, and the handler, then deployed zero-shot on a modified quadruped. At runtime, onboard RGB images, lidar scans, odometry, and handle-encoder measurements are fused into the same 200x200 BEV representation, and the trained policy outputs forward, lateral, and yaw commands, with a reactive lidar safety filter as a final layer. In controlled real-world runs, the policy followed paths and rounded obstacles while keeping the handler out of dan

What carries the argument

The BEV semantic map is the load-bearing object: a top-down grid showing walkable floor or sidewalk, walls and obstacles from accumulated lidar voxels, detected objects and pedestrians, and the robot and handler as circular regions. The same representation is rendered in simulation for reinforcement-learning training and reconstructed onboard from real sensors, so the policy sees matched inputs at deployment. The policy is a small convolutional network trained with a standard on-policy RL algorithm, and a local lidar safety filter blocks commands that would drive into nearby geometry.

Load-bearing premise

That the bird's-eye-view images produced in real time from onboard camera and lidar are close enough to the simulator's training images that the reinforcement-learned policy's collision-avoidance behavior carries over without any retuning.

What would settle it

A controlled experiment with at least 20 rollouts on a fixed obstacle course, counting robot and handler collisions and measuring minimum clearance, would settle the claim: if the learned policy does not reduce handler collisions compared with the costmap baseline, or if it incurs robot collisions at ordinary walking speeds, the central claim of safer collaborative navigation fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If Milo works as claimed, a roughly $2,000 open-source robot can take over the local, moment-to-moment guidance tasks — staying on path and avoiding obstacles — that today require dogs costing tens of thousands of dollars.
  • Because the robot needs no prior map and no external compute, it can be deployed in homes, sidewalks, and indoor facilities the moment it is turned on.
  • The handle-mounted magnetic encoders give the robot a live estimate of the handler's relative position, so the policy optimizes for the handler's body rather than only the robot's footprint, a direct improvement over costmap planners that ignore the handler.
  • The policy trains in about 10 minutes of GPU time, making the approach cheap enough to reproduce and iterate on, and the open release includes hardware instructions and software.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the BEV abstraction implies that future sim-to-real work on assistive navigation could concentrate on making the onboard perception match the simulator's rendering, since the control policy already transferred zero-shot in the reported runs.
  • Editorial inference: the sensorized two-degree-of-freedom handle could be extended from passive sensing to haptic feedback, letting the robot communicate turns or hazards through the handle itself.
  • Editorial inference: the same BEV stack could be pushed toward street crossings if traffic-light and vehicle detection were added, which the paper lists as future work rather than a demonstrated capability.
  • Editorial inference: the $2,000 cost is for the research platform; a production device would still need certification, durability, and support, so the accessible price is not yet settled.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper presents Milo, an open-source robotic guide dog built on a Unitree Go2 quadruped, a Jetson Orin Nano, a custom two-DoF handle with magnetic encoders, and a perception stack (LiDAR voxel mapping, NanoSAM walkable-path segmentation, YOLO detection, and Depth Anything pedestrian depth). A top-down BEV representation is fed to an RL policy trained in a Taichi-based 2D simulator; the trained policy is deployed zero-shot on the robot. The system is evaluated in three real-world scenarios—path following, static obstacle avoidance, and pedestrian avoidance—against a costmap baseline, with claims of smoother navigation and fewer handler collisions. The paper's central claim is that Milo is the first open-source, low-cost (~$2k USD) robotic guide dog capable of fully autonomous indoor/outdoor collaborative navigation.

Significance. If the central claim is established, Milo would be a valuable contribution: it demonstrates a fully onboard, map-free, handler-aware guide-dog platform at a fraction of the cost of a trained animal, with an open-source hardware/software release. The detailed appendices (simulator, RL reward design, safety filter, perception evaluation) are a strength, as is the fast training time. However, the empirical evidence is currently too thin to support the strength of the claims; the evaluation consists of three rollouts per condition, includes robot collisions, and abandons the pedestrian experiment as non-repeatable. The sim-to-real transfer assumption—that a policy trained on clean synthetic BEV frames survives real perception errors—is not demonstrated. These are evidence-strength concerns, not internal inconsistencies.

major comments (4)
  1. [§4.2–4.4, App. D] The zero-shot sim-to-real transfer is not established. The simulator (App. D) renders clean semantic masks with only Perlin noise on the road, while real BEV inputs contain NanoSAM/YOLO/Depth Anything errors, LiDAR blind spots, and 4 Hz temporal dropouts. Appendix A documents that NanoSAM mis-segments a wall as walkable when facing a wall. The evaluation is only 3 rollouts per scenario; the policy incurred robot collisions in 2 path-following runs at higher speeds, contacted the planter in 2/3 obstacle runs, and the pedestrian experiment was abandoned as non-repeatable (§4.4). This does not support the claim that policy behavior transfers to real perception noise.
  2. [Abstract, §1, §4.2] The claim of 'smoother navigation' is not supported by any quantitative metric. The experiments report completion times and collision counts only; there is no measurement of path smoothness, jerk, acceleration, or clearance variance. The reward function includes a jerk penalty (App. E, Eq. 8), but no corresponding evaluation metric is reported. The comparative claim vs. the costmap baseline rests on this unsupported qualitative descriptor.
  3. [§4.1, §4.2] The costmap baseline is deliberately blind to the handler ('does not have any awareness of the handler's relative position', §4.2). Consequently, the result that Milo has fewer handler collisions is largely by construction: the baseline cannot avoid handler collisions. This comparison is best framed as an ablation of handler-awareness, not as evidence that the learned policy is superior to a competitive navigation baseline. The paper should temper the comparative language or add a handler-aware baseline.
  4. [§4.4, Abstract] The pedestrian-avoidance claim in the abstract and introduction is not supported by the reported experiments. §4.4 states that the planned pedestrian course was abandoned due to 'little repeatability,' and the only evidence is qualitative observation and a supplementary video. Given that pedestrian avoidance is one of the three core capabilities claimed, this is a load-bearing gap. The authors should either provide a repeatable quantitative pedestrian evaluation or revise the scope of the claim.
minor comments (6)
  1. [§3.3] Typo: 'V oxels' should be 'Voxels'. Also, the LiDAR voxel resolution (10 cm) is mentioned in §4.1 but should be stated in §3.3 where the LiDAR processing is described.
  2. [§4.2] The costmap and policy runs are at different speeds, making the completion-time comparison difficult to interpret; report speed-matched runs or discuss the trade-off explicitly.
  3. [App. D] Duplicate phrase: 'the robot is spawned generally spawned towards the edge' — likely 'the robot is generally spawned towards the edge'.
  4. [App. D] The viewport size (8 m vs 4 m) is a significant design parameter; the parenthetical explanation of choosing 4 m due to LiDAR reliability should be moved to the main text or experimental setup, as it directly affects the observation space.
  5. [§6] The limitations section does not mention the absence of statistical testing or the small number of rollouts; consider adding an explicit statement about evaluation breadth and the abandoned pedestrian experiment.
  6. [Table 1] The comparison omits recent works such as [7] and [12] that could be relevant; if they are excluded because they do not meet the listed criteria, consider adding a footnote explaining the exclusion.

Circularity Check

0 steps flagged

No circularity found: the simulator-trained policy is evaluated against a separate real-world perception stack, with no load-bearing self-citations or definitional reductions.

full rationale

Milo's central claim is an empirical sim-to-real system result, not a derivation that reduces to its inputs. The navigation policy is trained in a Taichi BEV simulator with randomized layouts and reward terms, while the real-world evaluation uses an independently constructed perception stack (NanoSAM, YOLO, Depth Anything, LiDAR voxel mapping); the policy is deployed zero-shot and measured on real collisions, completion times, and path-following. The potential suspicion that the 'fewer handler collisions' result is forced by construction—because the costmap baseline lacks handler awareness while the policy reward includes handler-collision penalties—is an experimental-design fairness concern, not a definitional equivalence: the sim-to-real gap makes the outcome contingent, and the paper does not claim to predict real-world performance from the reward coefficients. The paper contains no load-bearing self-citations; references such as Sorokin, Rudin, and BEVFusion are external prior work. Appendices explicitly document limitations (NanoSAM wall mis-segmentation, non-repeatable pedestrian trials, 4 m viewport chosen due to LiDAR reliability), which further supports that the work is not hiding a circular dependency behind an unstated assumption. Accordingly, no circular step meets the evidentiary bar, and the score is 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The free parameters are all in the simulator/reward design and are tuned manually in simulation or on the robot design loop; they are not fitted to the real-world evaluation metrics, so they are not circular with the reported results, but they do weaken the transfer guarantee. The axioms are domain assumptions about perception reliability, sim-to-real fidelity, flat ground, handler shape, and LiDAR filtering, each of which is load-bearing for the autonomy claims.

free parameters (6)
  • r_inner (inner safety-zone radius) = 0.6 m
    Chosen empirically in App. E: smaller values led to more aggressive/dangerous policies, larger values to skittish ones.
  • r_outer (outer safety-zone radius) = 1.2 m
    Chosen empirically in App. E, same tuning rationale as r_inner.
  • v_fwd_target (target forward velocity) = 0.9 m/s
    Set empirically by walking-speed comfort (App. E).
  • sigma (velocity reward smoothness) = 0.3
    Found empirically (App. E).
  • Reward coefficients (10 terms) = -4.0, -3.0, -2.0, -0.1, 2.0, -0.1, -0.1, -0.1, -2.0, 1.0
    All reward weights were 'determined empirically based on success rate in the simulated environment' (App. E, Table 3).
  • BEV viewport size = 4 m local area
    Chosen empirically because the Go2 L2 LiDAR is not reliable at the desired 8m viewport (App. D footnote).
axioms (5)
  • domain assumption NanoSAM prompt-based segmentation with prompt points near the lower image center yields the correct walkable region, which requires the robot to start on or near a walkable path and face it.
    Invoked in Section 3.3 and Appendix A; the paper itself notes (Appendix A, last row of Fig. 7) that when facing a wall, NanoSAM mislabels the wall as walkable; acknowledged as a limitation in Section 6.
  • domain assumption The simulated BEV distribution (roads with Perlin-noise segmentation artifacts, walls, handlers, pedestrians) matches the real-world perception outputs closely enough for zero-shot policy transfer.
    The policy trained in the Taichi simulator is deployed directly on the robot without fine-tuning (Section 1); the assumption is only tested in a few rollouts, with known failure cases.
  • domain assumption The ground is flat for the homography-based projection of walkable-path masks onto the BEV plane.
    Ground projection in Section 4.1; Section 6 states 'the current ground-projection method assumes flat terrain, and handling uneven surfaces such as slopes and ramps remains future work.'
  • domain assumption The handler's state can be represented as a fixed-radius circle in the BEV map, derived from handle azimuth/elevation encoders, ignoring forces, leaning, or non-circular footprints.
    Section 3.3: 'This position is represented as a circular region in the BEV map'; the simulation also draws the handler as a circle (App. D).
  • domain assumption LiDAR height-threshold voxel filtering (dropping voxels below/above height bounds) removes ground and overhanging ceilings/tree canopies without discarding true obstacles.
    Section 3.3 states voxels below a height threshold are removed to filter ground and voxels above a maximum height are discarded, assuming the robot and handler can safely pass under such structures.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Milo, a Fully Autonomous Indoor/Outdoor Robotic Guide Dog." pith.science (2026). https://pith.science/paper/BE35ULCO

@misc{pith2026260719530,
  author       = {Pith},
  title        = {Pith review of: Milo, a Fully Autonomous Indoor/Outdoor Robotic Guide Dog},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BE35ULCO}},
  note         = {Machine review of arXiv:2607.19530}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Many Blind and Low-Vision (BLV) people rely on guide dogs for moment-to-moment navigation, such as staying on path and avoiding obstacles and pedestrians. However, guide dogs are expensive to acquire and maintain (approximately \$50k USD plus ongoing costs), often involve long waiting lists, and have relatively short life expectancies. While robot guide dogs offer a promising alternative, existing approaches exploring this idea suffer from several drawbacks: They often lack the autonomy required for real-world deployment, relying on prior 3D scans of the environment, external computation, or limited awareness of the handler. In this work, we present Milo, the first open-source, low-cost (approximately \$2k USD) robotic guide dog platform capable of fulfilling the basic collaborative navigation role expected of a guide dog. Milo is fully autonomous, requiring no a priori knowledge of the environment, completely self-contained with all computation performed onboard, and suitable for both indoor and outdoor navigation while avoiding obstacles and pedestrians. Our system consists of a modified Unitree Go2 robot (equipped with onboard compute, sensors, and a handle), a perception stack combining voxel mapping with floor, obstacle, and pedestrian detection, and a navigation stack based on an obstacle-avoidance policy trained in a custom bird's-eye-view simulator. We evaluate Milo in real indoor and outdoor obstacle courses and compare it against a costmap-based baseline, demonstrating smoother navigation and fewer handler collisions. To maximize accessibility for BLV users, we release both the robot hardware instructions and the complete software stack as open source.

Figures

Figures reproduced from arXiv: 2607.19530 by Christopher Pal, Florian Golemo, Joanna Wolski, Joel Ruben Antony Moniz.

Figure 1
Figure 1. Figure 1: Milo autonomously guides a handler along indoor and outdoor paths while avoiding ob￾stacles and oncoming pedestrians, adapting to the handler’s position, desired walking speed, and turn commands. The project is released as open source. Abstract: Many Blind and Low-Vision (BLV) people rely on guide dogs for moment-to￾moment navigation, such as staying on path and avoiding obstacles and pedestri￾ans. However… view at source ↗
Figure 3
Figure 3. Figure 3: Method Overview. (a) RGB images, LiDAR scans, odometry, and handle encoder mea￾surements are acquired. (b) RGB images are processed to segment the walkable path and detect objects and pedestrians, while depth estimation updates pedestrian positions. (c) The segmented path, detected objects and pedestrians, the robot and handler states, and LiDAR data are fused into a voxel-based 3D map. (d) The voxel map i… view at source ↗
Figure 2
Figure 2. Figure 2: Hardware Overview. (A) The Nvidia Jetson Orin Nano runs navigation and percep￾tion systems; (B) the 2 degree-of-freedom joint is equipped with magnetic encoders that sense the handler’s position with respect to the dog; (C) the telescopic handle is adjustable to cater to different handler heights and preferences; (D) with the 5- way directional pad, the handler can start/stop the robot, indicate faster/slo… view at source ↗
Figure 4
Figure 4. Figure 4: Example of Costmap-Based Tra￾jectory Selection. The segmented camera frame is projected onto the ground plane to generate trajectories. The center-right trajec￾tory is selected as it has the lowest cost; ties are resolved in favor of the rightmost trajec￾tory. Costmap Baseline. As a baseline for evaluating our learned navigation policy, we implemented a classic costmap-based controller similar to Hwang et … view at source ↗
Figure 5
Figure 5. Figure 5: Experiment - Path Following. We compared the costmap controller (cyan) at fixed speed to the policy (pink) at variable speed. Starting at the left, both approaches had to cross the bridge and were timed at the start and finish. The costmap led the handler out of bounds for several meters (start marked on the map with X). The policy incurred 2 robot collisions (marked with X) at higher speeds but otherwise … view at source ↗
Figure 6
Figure 6. Figure 6: Experiment - Obstacle Avoidance. Milo and the handler started at the top right facing left and navigated around a rubbish bin and a planter. All costmap runs resulted in handler collisions with the bin, while the policy runs protected the handler, al￾though two grazed the planter. We also tested turning around static obstacles, for which we set up a relatively tight turn for the robot & handler unit. We ma… view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of NanoSAM and SegFormer walkable-surface segmentation. Columns show the input RGB image, NanoSAM prediction, and SegFormer prediction, respectively. Rows correspond to an indoor floor, an outdoor sidewalk, a tiled pedestrian area, a park path, a sidewalk partially covered by snow, and a camera frame where the robot is facing a wall. Both methods perform well on standard indoor-floor and outdoor… view at source ↗
Figure 8
Figure 8. Figure 8: Estimated distance to a pedestrian as a function of time for several camera-based localization approaches. All methods were evaluated on the same video recording of a person walking toward the robot. Ground-truth distances were obtained using an AirTag carried by the pedestrian and were interpolated to produce a smooth reference trajectory. The figure compares three methods for estimating the distance of p… view at source ↗
Figure 9
Figure 9. Figure 9: Example Simulator Frame. Visualization of the simulated environment. The traversable road is shown in grey (with intentional gaps to mimic imperfect floor segmentation), walls in red, handler in purple, pedestrian in yellow, and obstacle in blue. Our birds-eye-view simulator, implemented in Taichi [9], constructs scenes like this: Roads are random lines and Bezier curves [20] drawn in OpenCV that can be ro… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 7 linked inside Pith

  1. [1]

    R. R. Bourne, J. D. Steinmetz, S. Flaxman, et al. Trends in prevalence of blindness and dis- tance and near vision impairment over 30 years: an analysis for the global burden of dis- ease study.The Lancet Global Health, 9(2):e130–e143, 2021. ISSN 2214-109X. doi:10. 1016/S2214-109X(20)30425-3. URLhttps://www.thelancet.com/journals/langlo/ article/PIIS2214-...

    2021

  2. [2]

    Guide the way, 2025

    CNIB Foundation. Guide the way, 2025. URLhttps://www.guidedog.org/gd/ about-us/about-the-guide-dog-foundation.aspx. Accessed: 2026-05-28

    2025

  3. [3]

    Sorokin, J

    M. Sorokin, J. Tan, C. K. Liu, and S. Ha. Learning to navigate sidewalks in outdoor environ- ments.IEEE Robotics and Automation Letters, 7(2):3906–3913, 2022

    2022

  4. [4]

    S. Cai, A. Ram, Z. Gou, M. A. W. Shaikh, Y .-A. Chen, Y . Wan, K. Hara, S. Zhao, and D. Hsu. Navigating real-world challenges: A quadruped robot guiding system for visually impaired people in diverse environments. InProceedings of the 2024 CHI Conference on Human Factors in Computing Systems, pages 1–18, 2024

    2024

  5. [5]

    Hwang, T

    H. Hwang, T. Xia, I. Keita, K. Suzuki, J. Biswas, S. I. Lee, and D. Kim. System configuration and navigation of a guide dog robot: Toward animal guide dog-level guiding work.arXiv preprint arXiv:2210.13368, 2022

    Pith/arXiv arXiv 2022

  6. [6]

    Hwang, S

    H. Hwang, S. Yang, J. S. Monon, N. A. Giudice, S. I. Lee, J. Biswas, and D. Kim. GuideNav: User-informed development of a vision-only robotic navigation assistant for blind travelers,

  7. [7]

    Hayamizu, D

    Y . Hayamizu, D. DeFazio, H. Mehta, Z. Altaweel, J. Choe, C. Lin, J. Juettner, F. Xiao, J. Black- burn, and S. Zhang. From woofs to words: Towards intelligent robotic guide dogs with verbal communication, 2026. URLhttp://arxiv.org/abs/2603.12574

    arXiv 2026

  8. [8]

    Hwang, H.-T

    H. Hwang, H.-T. Jung, N. A. Giudice, J. Biswas, S. I. Lee, and D. Kim. Towards robotic com- panions: Understanding handler-guide dog interactions for informed guide dog robot design,

  9. [9]

    Hu, T.-M

    Y . Hu, T.-M. Li, L. Anderson, J. Ragan-Kelley, and F. Durand. Taichi: a language for high- performance computation on spatially sparse data structures.ACM Transactions on Graphics (TOG), 38(6):201, 2019

    2019

  10. [10]

    Chen and B

    J. Chen and B. Zhang. Exploration and navigation in unknown environments for guide dog robots. In2025 9th International Conference on Robotics and Automation Sciences (ICRAS), pages 48–326, 2025. doi:10.1109/ICRAS65818.2025.11108808. URLhttps: //ieeexplore.ieee.org/document/11108808. ISSN: 2694-3506

    arXiv 2025

  11. [11]

    Viteri and C.-H

    J. Viteri and C.-H. G. Li. Autonomous sidewalk navigation featuring end-to-end RGB-d dual- ConvNet steering. In2024 IEEE International Conference on Advanced Intelligent Mechatron- ics (AIM), pages 703–708. IEEE, 2024. ISBN 979-8-3503-5536-9. doi:10.1109/AIM55361. 2024.10637141. URLhttps://ieeexplore.ieee.org/document/10637141/. 10

    arXiv 2024

  12. [12]

    J. T. Kim, W. Yu, Y . Kothari, J. Tan, G. Turk, and S. Ha. Transforming a quadruped into a guide robot for the visually impaired: Formalizing wayfinding, interaction modeling, and safety mechanism.arXiv preprint arXiv:2306.14055, 2023

    Pith/arXiv arXiv 2023

  13. [13]

    Zhang, D

    C. Zhang, D. Han, Y . Qiao, J. U. Kim, S.-H. Bae, S. Lee, and C. S. Hong. Faster segment anything: Towards lightweight sam for mobile applications.arXiv preprint arXiv:2306.14289, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    Nanosam.https://www.jetson-ai-lab.com/archive/vit/tutorial_ nanosam.html, 2024

    NVIDIA. Nanosam.https://www.jetson-ai-lab.com/archive/vit/tutorial_ nanosam.html, 2024. Accessed: 2026-05-28

    2024

  15. [15]

    Jocher and J

    G. Jocher and J. Qiu. Ultralytics yolo26, 2026. URLhttps://github.com/ultralytics/ ultralytics

    2026

  16. [16]

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao. Depth anything v2. arXiv:2406.09414, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    Rudin, D

    N. Rudin, D. Hoeller, P. Reist, and M. Hutter. Learning to walk in minutes using massively parallel deep reinforcement learning. arxiv.arXiv preprint arXiv:2109.11978, 2021

    Pith/arXiv arXiv 2021

  18. [18]

    Liang, H

    T. Liang, H. Xie, K. Yu, Z. Xia, Z. Lin, Y . Wang, T. Tang, B. Wang, and Z. Tang. Bevfusion: A simple and robust lidar-camera fusion framework.Advances in Neural Information Processing Systems, 35:10421–10434, 2022

    2022

  19. [19]

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo. Segformer: Simple and efficient design for semantic segmentation with transformers.CoRR, abs/2105.15203, 2021. URLhttps://arxiv.org/abs/2105.15203

    Pith/arXiv arXiv 2021

  20. [20]

    Prautzsch, W

    H. Prautzsch, W. Boehm, and M. Paluszny.B´ezier and B-spline techniques, volume 6. Springer, 2002

    2002

  21. [21]

    skittish

    K. Perlin. An image synthesizer.ACM SIGGRAPH Computer Graphics, 19(3):287–296, 1985. doi:10.1145/325165.325247. 11 Appendix A Evaluation of Walkable Surface Segmentation Methods Figure 7:Comparison of NanoSAM and SegFormer walkable-surface segmentation.Columns show the input RGB image, NanoSAM prediction, and SegFormer prediction, respectively. Rows corre...

    arXiv 1985

  22. [2024]

    URLhttp://arxiv.org/abs/2402.06790

    Pith/arXiv arXiv

  23. [2025]

    URLhttp://arxiv.org/abs/2512.06147

    arXiv

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.