RoboAtlas: Surgical Geometry & Active SLAM

• RoboAtlas is a dual-use concept encompassing an anatomical geometric atlas for otolaryngologic robot design and a contextual Active SLAM framework for autonomous navigation.
• The geometric component derives from CT scans to define mean dimensions, variability, and workspace limits critical for safe endoscopic tool design.
• The Active SLAM component uses hybrid mapping (TSDF, octree) and a contextual multi-armed bandit to balance exploration with semantic reasoning for real-time decision-making.

In the supplied arXiv literature, RoboAtlas denotes two distinct constructs. In otolaryngologic robotics, it refers to a geometric atlas of the middle ear and paranasal sinuses derived from CT measurements and expressed as mean dimensions, standard deviations, and 90% ranges for robotic workspace design (Michel et al., 2021). In embodied AI, it denotes a contextual Active SLAM framework that combines OpenRoboVox, frontier exploration, global semantic-map reasoning, and egocentric VLM reasoning through a contextual multi-armed bandit, with evaluation in simulation and on a Unitree Go2 robot (Schperberg et al., 24 Jun 2026). The shared label therefore spans two different technical programs: patient-specific workspace characterization for endoscopic robot design, and large-scale semantic navigation grounded in 3D mapping.

1. Distinct uses of the term in the literature

Usage	Core definition	arXiv id
Geometric RoboAtlas	Parametric description of linear dimensions, volumes, and angular limits for middle-ear and sinus workspaces	(Michel et al., 2021)
Active-SLAM RoboAtlas	Contextual Active SLAM framework using OpenRoboVox and a contextual multi-armed bandit	(Schperberg et al., 24 Jun 2026)

The distinction is substantive rather than nominal. The 2021 surgical work is an anatomical and design-oriented atlas: it defines workspace bounds for endoscopic robotic applications and explicitly does not construct a formal PCA-based Statistical Shape Model (Michel et al., 2021). The 2026 system is a robot autonomy stack: it treats RoboAtlas as a decision-making framework that balances geometric exploration and semantic reasoning in large-scale environments (Schperberg et al., 24 Jun 2026).

A common misconception is to treat both as instances of the same software or robotic platform. The available literature does not support that reading. One use is anatomical and preoperative-design oriented; the other is algorithmic and online-navigation oriented.

2. Geometric RoboAtlas for otolaryngologic robotic applications

The geometric RoboAtlas was introduced to help define the workspace for robots studied in otolaryngologic surgery, with the stated aim of helping design and optimize such robots while accounting for patient variability (Michel et al., 2021). The study used scans of several patients of different ages and sexes to determine average workspace size in two anatomically linked endoscopic domains: the middle ear and the paranasal sinuses.

The patient cohorts were modest but explicitly specified. The ear workspace cohort comprised 36 patients (17 right, 19 left), age 2–81 (mean 39 y), sex ratio 1:1. The sinus workspace cohort comprised 23 patients, age 11–95, with only one patient younger than 20 years. Exclusions included congenital malformations (e.g. aplasia) and prior surgery with major bone distortion. All measurements were performed on routine CT acquired in 2018 at Nantes University Hospital, using Vue PACS v11.3 (Carestream) (Michel et al., 2021).

Methodologically, the atlas is based on direct PACS measurements rather than volumetric registration. The paper reports that no volumetric mesh or point-cloud registration was performed and that no explicit coordinate system definition or Procrustes-style registration was reported. Ear landmarks included CAEd_lateral, CAEd_medial, CAElength, OMheight, OMwidth, and OMap_length. Sinus landmarks included Workspace_depth, Workspace_width, Nasal_fossae_width, Workspace_height, and Piriforme_orifice_height. For each measurement, the study computed mean $\mu$ and standard deviation $\sigma$ in Excel v.14.5, assessed normality by $r^2$ of normal-fit curves with $r^2 > 0.9$ for ear and $> 0.8$ for sinuses, and extracted 90% percentile ranges (Michel et al., 2021).

The paper is explicit about what RoboAtlas is not. It states that it does not construct a formal PCA-based Statistical Shape Model, and that no point-cloud alignment, covariance matrices or principal modes $\Phi$ were computed, nor shape-parameter vectors $b$ derived. In that sense, the atlas is a parametric anatomical summary rather than a generative shape model (Michel et al., 2021).

3. Workspace metrics, design constraints, and limitations of the geometric atlas

The atlas reports linear dimensions, volume proxies, and angular limits for two workspaces: the external auditory canal plus tympanic membrane region, and the paranasal sinus corridors (Michel et al., 2021).

Region	Measurement	Reported value
External auditory canal	CAEd_lateral	mean 6.1 mm, range 3.9–7.3 mm, $\sigma \approx 0.8$ mm
External auditory canal	CAEd_medial	mean 7.9 mm, range 6.2–11.1 mm, $\sigma \approx 1.1$ mm
External auditory canal	CAElength	mean 26.9 mm, range 22.5–35.3 mm, $\sigma \approx 3.2$ mm
External auditory canal	Canal volume	mean 1.32 cm³
Tympanic membrane workspace	OMheight	16.2 mm, range 14.1–19.4 mm, $\sigma$0 mm
Tympanic membrane workspace	OMwidth	10.8 mm, range 7.6–12.3 mm, $\sigma$1 mm
Tympanic membrane workspace	OMdepth	5.7 mm, range 3.1–7.2 mm, $\sigma$2 mm
Tympanic membrane workspace	Eardrum volume proxy	0.99 cm³
Paranasal sinuses	Workspace_depth	77.04 mm, range 59–94 mm, $\sigma$3 mm
Paranasal sinuses	Workspace_width	39.26 mm, range 27–47 mm, $\sigma$4 mm
Paranasal sinuses	Nasal_fossae_width	13.74 mm, range 9–18 mm, $\sigma$5 mm
Paranasal sinuses	Workspace_height	55.39 mm, range 42–67 mm, $\sigma$6 mm
Paranasal sinuses	Piriforme_orifice_height	29.57 mm, range 21–36 mm, $\sigma$7 mm

The 90% population bounds are central to the atlas’s engineering interpretation. For the ear, 90% of patients lay within CAEd_lateral $\sigma$8 mm, CAEd_medial $\sigma$9 mm, and CAElength $r^2$0 mm. For the tympanic membrane workspace, the corresponding 90% ranges were height $r^2$1 mm, width $r^2$2 mm, and depth $r^2$3 mm. For the sinuses, the paper reports Workspace_depth $r^2$4 mm, Nasal_fossae_width $r^2$5 mm, Workspace_height $r^2$6 mm, and Piriforme_orifice_height $r^2$7 mm for 90% of patients (Michel et al., 2021).

The atlas also includes entry-vector constraints. In the ear, the ideal remote center is described as being at canal entrance & tympanic membrane, allowing $r^2$8 of insertion angles in optimal head position, with up to $r^2$9 in extremes. In the sinuses, the pivot is placed at piriform aperture, with full max travel of endoscope pivot $r^2 > 0.9$0, and clinically $r^2 > 0.9$1 endoscopes ($r^2 > 0.9$2) (Michel et al., 2021).

These measurements are translated into robot-design guidelines. The paper recommends at least 35 mm of axial travel for the ear and up to 95 mm insertion depth for the sinuses. Angularly, it recommends $r^2 > 0.9$3 about the ear entry point to cover head-tilting extremes and $r^2 > 0.9$4 pivot for a $r^2 > 0.9$5 endoscope, with full mechanical allowance of $r^2 > 0.9$6–$r^2 > 0.9$7 in the sinus case. The reported smallest ear canal diameter of 3.9 mm leads to the requirement that robot/tool cross-section must be $r^2 > 0.9$8 mm (endoscope) + auxiliary tools ($r^2 > 0.9$9 mm cannula) with safety margin. The paper further specifies endoscope $> 0.8$0 mm, suction $> 0.8$1–$> 0.8$2 mm, micro-instruments $> 0.8$3–$> 0.8$4 mm, and recommends designing kinematic chains around these diameters with at least 0.5 mm clearance (Michel et al., 2021).

Several limitations are explicit. The study reports sample sizes modest (36 ears, 23 sinuses), under-representation of pediatric (< 10 y) and extreme morphologies, no statistically significant differences by side or gender in most measures, and slight age effects on canal diameters. Recommended future work includes increasing the cohort to > 50 per region, including younger children (< 10 y), constructing a true PCA-based statistical shape model of ear canal and sinus cavity surfaces, and extending the atlas to other endoscopic sites (e.g. larynx, skull-base cavities) (Michel et al., 2021).

4. RoboAtlas as contextual Active SLAM

The 2026 RoboAtlas framework addresses contextual Active SLAM by adaptively balancing geometric exploration and semantic reasoning using OpenRoboVox, described as a scalable 3D semantic mapping system (Schperberg et al., 24 Jun 2026). Its architecture couples metric reconstruction, instance-level semantic fusion, compact scene abstraction, and expert-based decision-making.

The map representation has both geometric and semantic components. Geometrically, OpenRoboVox maintains a Truncated Signed Distance Field (TSDF) running at ~15 Hz and an octree/OctoMap projected into a 2D occupancy grid $> 0.8$5 for navigation. Semantically, it uses a probabilistic voxel grid in which each voxel $> 0.8$6 maintains a Dirichlet posterior $> 0.8$7 over a growing set of global instance IDs $> 0.8$8. Instance labels are produced per frame by YOLO-World for bounding boxes and the Tokenize Anything model (TAP) for segmentation masks; the resulting 2D detections are back-projected into the voxel grid and fused via Bayesian updates (Schperberg et al., 24 Jun 2026).

To avoid repeated full-map scans, a background thread maintains a compact Scene-Dictionary $> 0.8$9 of object instances. On each cycle, it selects a small batch $\Phi$0 via a priority union of temporal recency $\Phi$1, active observation $\Phi$2, and rotational consistency $\Phi$3. The dictionary is updated incrementally as

$\Phi$4

where $\Phi$5 computes an instance’s centroid, bounding-box, and occupancy probability from its voxels. The system also includes a 2D Pillar Map, which projects the 3D voxel set $\Phi$6 onto a 2D grid $\Phi$7 at resolution $\Phi$8, storing in each cell the top-K MAP-assigned instances ranked by voxel count $\Phi$9 (Schperberg et al., 24 Jun 2026).

The implementation emphasizes scalability. Volumetric map expansion is performed in small cubic blocks $b$0 to bound peak VRAM,

$b$1

and the geometric (15 Hz) and semantic (≈5 Hz) pipelines run asynchronously on separate threads, with fine-grained locks only at fusion points (Schperberg et al., 24 Jun 2026).

5. Contextual multi-armed bandit, semantic reasoning, and VLM experts

RoboAtlas formalizes decision-making as a contextual multi-armed bandit (CMAB) with three experts, or arms: frontier exploration $b$2, semantic-map reasoning $b$3, and egocentric VLM reasoning $b$4 (Schperberg et al., 24 Jun 2026).

The context vector is

$b$5

where $b$6 is the fraction of grid covered, $b$7 is the change in coverage since the last step, $b$8 is a backtracking indicator, $b$9 is the previous arm, $\sigma \approx 0.8$0 is a confidence score from the egocentric VLM, and $\sigma \approx 0.8$1 is a soft count of candidate observations above cosine-similarity thresholds $\sigma \approx 0.8$2,

$\sigma \approx 0.8$3

The bandit uses LinUCB (Disjoint UCB). For each arm $\sigma \approx 0.8$4, the framework maintains $\sigma \approx 0.8$5 and $\sigma \approx 0.8$6, computes

$\sigma \approx 0.8$7

selects $\sigma \approx 0.8$8, and updates $\sigma \approx 0.8$9 and $\sigma \approx 1.1$0 using the observed reward (Schperberg et al., 24 Jun 2026).

The reward is defined as

$\sigma \approx 1.1$1

with $\sigma \approx 1.1$2, $\sigma \approx 1.1$3, $\sigma \approx 1.1$4, $\sigma \approx 1.1$5, and $\sigma \approx 1.1$6. This formulation explicitly combines geometric exploration signals, anti-backtracking pressure, VLM confidence, semantic similarity, and terminal success (Schperberg et al., 24 Jun 2026).

Each expert has a distinct operational role. The frontier expert clusters frontier cells into representative centers $\sigma \approx 1.1$7, computes information gain $\sigma \approx 1.1$8, evaluates path length $\sigma \approx 1.1$9 with A*, checks energy feasibility, and selects

$\sigma \approx 3.2$0

The semantic-map expert filters the Scene-Dictionary using precomputed 2D potential fields $\sigma \approx 3.2$1 and neighbor similarity $\sigma \approx 3.2$2, serializes top-relevant instances into a token-budgeted prompt containing each instance’s caption, centroid $\sigma \approx 3.2$3, bounding box $\sigma \approx 3.2$4, and similarity vector $\sigma \approx 3.2$5, then queries GPT-4o or Qwen2.5-VL for a JSON ranking. The egocentric VLM expert takes current RGB and depth with spatial annotations on a coarse image grid, queries GPT-4o or Qwen2.5-VL, and returns both per-annotation confidence $\sigma \approx 3.2$6 and a discrete action in {“go-to”, “turn-left”, “turn-right”, “about-face”}, rolling out up to $\sigma \approx 3.2$7 steps before resubmitting when a turn action is selected (Schperberg et al., 24 Jun 2026).

A second common misconception is that the system merely substitutes a large language or vision-LLM for classical exploration. The reported architecture is instead explicitly hybrid: the VLM-based experts are grounded in a large-scale semantic map, and semantic as well as VLM inference are asynchronous so that neither blocks the real-time SLAM thread (Schperberg et al., 24 Jun 2026).

6. Empirical performance, ablations, and stated limitations

RoboAtlas was evaluated in Isaac Sim, Habitat (HM3D Val-Unseen, 36 scenes, 278 subtasks), and on real hardware: Unitree Go2 quadruped + Jetson Orin + WiFi→desktop RTX 4090 (Schperberg et al., 24 Jun 2026). The OpenRoboVox scale test on hardware mapped 1 803 m² over two floors and recorded 29 588 unique instances in one run.

Evaluation	Configuration	Reported result
Isaac Sim CMAB ablation	Frontier only	SR = 53.3%, SPL = 0.386
Isaac Sim CMAB ablation	Ego-VLM only	SR = 66.7%, SPL = 0.510
Isaac Sim CMAB ablation	Semantic-map only (cold)	SR = 0%
Isaac Sim CMAB ablation	RoboAtlas (CMAB)	SR = 100%, SPL = 0.920, average time 780 s
Hardware directives	6 varied language tasks	100% success, mean standoff 0.22 m, $\sigma \approx 3.2$8 m
GOAT-Bench Val-Unseen	Prior SOTA (HIMM with GPT-4o)	SR = 72.8%, SPL = 56.1%
GOAT-Bench Val-Unseen	RoboAtlas + GPT-4o	SR = 90.6%, SPL = 53.4%
GOAT-Bench Val-Unseen	RoboAtlas + Qwen2.5-VL	SR = 88.8%, SPL = 53.1%

On GOAT-Bench “Val Unseen”, RoboAtlas reports the highest reported success rate (SR) of 90.6% using GPT-4o, improving over the strongest prior baseline by 17.8 percentage points in SR. With Qwen2.5-VL-7B, it reports 88.8% SR, which the paper states outperforms all baselines using GPT-4o in SR. Per modality with GPT-4o, the reported SR values are 94.9% for object queries, 84.6% for descriptions, and 92.0% for image-based queries (Schperberg et al., 24 Jun 2026).

Bandit telemetry on GOAT-Bench is also reported: across 2 005 decisions, the picks were SemanticMap 38.7% (46.8% of path length), Frontier 16.4% (22.6%), and EgoVLM 45.0% (30.6%). The paper interprets this as stage-dependent allocation: early frontier, then VLM, then semantic exploitation (Schperberg et al., 24 Jun 2026).

The stated limitations are concrete. Semantic mapping errors, including missed detections or association noise, can propagate into planning. Reliance on remote API calls (GPT-4o) introduces latency and requires reliable network connectivity. The reward function is hand-tuned, and the paper suggests that a learned or more principled formulation could improve generalization. It also notes a slight SPL penalty in long episodes, describing this as an Active SLAM trade-off that could be tuned per task (Schperberg et al., 24 Jun 2026).

7. Relation to Atlas humanoid research and broader methodological context

RoboAtlas should be distinguished from a broader body of work on the Boston Dynamics Atlas humanoid, although there are methodological continuities. Atlas research has addressed partial footholds and line contacts through foothold exploration, CoP shifting, polygon cropping, and a whole-body momentum-based control algorithm running at 1 kHz, with continuous sequences of 4–6 steps on straight line-field in real tests and CoP tracking error ≈ 2 cm RMS (Wiedebach et al., 2016). It has also addressed large step-ups through a nonlinear trajectory optimization based on simplified centroidal dynamics, solved with IPOPT, reporting that peak torque dropped to ~96 Nm on the 0.35 m step (a 20% reduction) and from ~130 Nm down to ~105 Nm (19% reduction) on a 0.40 m step (Dafarra et al., 2020).

A separate Atlas navigation line introduced a two-stage rough-terrain planner using a 2.5D height map, A* with a traversability-aware edge cost,

$\sigma \approx 3.2$9

followed by gradient-descent path smoothing, and demonstrated the integrated pipeline on a DRC Boston Dynamics Atlas robot with 9/10 trials on cinder blocks and 8/10 on stones (McCrory et al., 2022). This suggests a broader methodological continuity in Atlas-related robotics: explicit environment representation, optimization or search over physically meaningful costs, and close integration with whole-body balance control. That continuity is suggestive rather than nominal, because the 2026 RoboAtlas framework was evaluated on a Unitree Go2 quadruped, not on Atlas (Schperberg et al., 24 Jun 2026).

Viewed across these works, the name RoboAtlas marks two unrelated but technically rigorous uses. In one, it is a patient-variability-aware anatomical parameterization for designing endoscopic surgical robots (Michel et al., 2021). In the other, it is a contextual Active SLAM architecture in which large-scale 3D semantic maps ground frontier exploration, LLM-based global reasoning, and egocentric VLM guidance (Schperberg et al., 24 Jun 2026). The coexistence of these meanings is unusual, but the underlying ambition is comparable: to replace underspecified operating assumptions with explicit, quantitatively constrained representations of the workspace.