Socially Compliant Navigation Dataset (SCAND): A Large-Scale Dataset of Demonstrations for Social Navigation
1. Paper overview
In one sentence: SCAND turns natural robot teleoperation into supervision for social navigation, with promising Spot baselines whose evidence remains limited to path imitation and simple participant-rated encounters. e02e03e04e05e12e13e14
| At a glance | What to know |
|---|---|
| Research problem | Source description Social navigation requires responding considerately to other agents, beyond simply avoiding obstacles. Explicit reward functions struggle to capture these contextual conventions. The authors identify a shortage of natural robot demonstrations pairing sensory observations with socially intended motion commands: perception datasets, scripted robot paths and simulated social experts serve different purposes. SCAND addresses this supervision gap through teleoperation in populated indoor and outdoor campus spaces. e02e03e06 |
| Core mechanism | |
| A key reported result | Perceived social compliance during physical Spot navigation: BC: mean 4.39, SD 0.99 Participant rating from 1 to 5; mean and standard deviation. Fourteen indoor participants; static obstacle-person and dynamic passing scenarios; algorithm order randomized; ratings summarized across scenarios. move_base: mean 2.86, SD 0.82 Participants rated BC more considerate. The authors report one-way ANOVA p < 0.001. The reported social-compliance mean of 4.39 is retained, but it is not reconstructed by a simple equal-weight average of Figure 7's approximately 4.85 static and 4.28 dynamic BC bars. The paper provides no raw ratings or pooling details to reconcile this discrepancy. The displayed preference for BC remains; the comparator is not designed for social compliance, and this is a small subjective study. e12e13 |
| Reading caution | Source description Collection in one city can encode regional conventions, including keeping right and overtaking on the left. Four demonstrators and campus-heavy demographics limit diversity; rare interactions remain difficult. The authors explicitly leave broader benchmarking and more sophisticated social scenarios for future work. e03e14 |
Core contributions
Figure 3. The collection shares core sensing while retaining information specific to wheeled and legged robots. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the labels to separate shared hardware from robot-specific measurements. Both platforms show a front RGB camera and Velodyne puck. Jackal adds stereo imagery and wheel odometry; Spot adds body cameras, visual odometry and leg joint angles. The small top panels identify Spot's right, rear and left cameras, while the lower label identifies its two front body cameras. Section III-A also specifies inertial sensing, recording rates and transforms between each sensor frame and the robot body. These details matter because a demonstration is a sequence of observations aligned with joystick commands, not simply the photographed robot or a path drawn through space. e04e09e16
What it supports. SCAND can support investigations of how morphology and sensing shape navigation demonstrations. Its rosbag streams and frame transforms provide the ingredients for constructing observation-action examples. However, the paper's learning analyses select Spot data; the breadth of the collection exceeds the demonstrated transfer scope of its baselines.
Where the evidence stops. The photograph inventories available sensing, not the inputs to every trained network. The reported BC baseline uses lidar, position and inertial history with planner priors. The paper does not demonstrate a policy trained across both robots or transfer between them.
2. Motivation
2.1 The problem and the proposed response
Social navigation requires responding considerately to other agents, beyond simply avoiding obstacles. Explicit reward functions struggle to capture these contextual conventions. The authors identify a shortage of natural robot demonstrations pairing sensory observations with socially intended motion commands: perception datasets, scripted robot paths and simulated social experts serve different purposes. SCAND addresses this supervision gap through teleoperation in populated indoor and outdoor campus spaces. e02e03e06
2.2 What this reading follows
A robot can avoid an obstacle yet still make a pedestrian uncomfortable. SCAND approaches that gap by recording how people choose to drive robots through shared spaces. Its 138 demonstrations combine motion commands with sensing on a wheeled Jackal and a legged Spot, plus labels describing encountered social situations. Read the paper in three layers: what the collection actually covers, how recorded behavior becomes a learning target, and what the validation measures. The strongest physical evidence is a fourteen-participant comparison with move_base. That evidence concerns perceived consideration and safety in two indoor scenarios; broader social competence remains an open evaluation problem. e02e03e04e05e12e13e14
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Datasets |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded dataset classification is supported by the multimodal navigation demonstrations and social tags. Architecture, prediction paradigm and world-action quadrant are appropriately not applicable to the resource itself. Its illustrative BC baseline shares representations across path and command heads, but predicts no environmental future; joint training alone does not make SCAND a unified world-action model. e02e04e05e09e10
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
5. Method in detail
5.1 Start with the supervision that teleoperation creates
SCAND's basic learning resource is the association between a robot's observations and the human commands that produced its motion. The operator follows the robot through naturally occurring interactions rather than constructing each encounter in a laboratory. Lidar, inertial measurements, odometry and camera streams describe what was sensed; joystick commands expose the chosen response. Sensor-to-body transforms allow those streams to be interpreted in robot coordinates. The trajectory tag list then identifies situations such as overtaking or waiting at a doorway. It does not replace the action sequence with a dense social-success label. This distinction explains why the dataset is suited to imitation learning while still leaving evaluation difficult: human demonstrations supply examples of behavior, but the paper does not provide an independently annotated standard for every socially acceptable alternative. e02e03e04e05e14
Figure 4. This diagram classifies who drove a recorded segment; the BC planner is a related architecture described in the caption and text. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow both branches from left to right. The upper branch encodes ten channels of BEV lidar with a four-layer CNN; the lower branch encodes joystick, recorded future trajectory and inertial information with dense layers. Their arrows converge on the classifier head and end at a demonstrator label. Section IV-A additionally includes relative robot positions. The future trajectory is therefore an input to retrospective classification, not an output or a predicted environmental state. The caption explicitly says the BC agent is not shown: that model changes the temporal window and inputs and substitutes global-path and local-command heads for the classifier. e07e08e09
What it supports. On sixteen trajectories from two demonstrators following the same route, the paper reports 74.48% held-out accuracy against a stated 50% guessing baseline. This supports distinguishable recorded behavior. The architecture also explains why that result cannot be read as an online social-navigation success rate.
Where the evidence stops. The figure, caption and text agree on the classifier's information flow. It does not depict the complete BC model. The classification study controls the route but does not isolate social style from all session, crowd or motion cues; the validation/test terminology also varies within the source.
5.2 Separate knowing who drove from learning what to drive
The classifier can look at a recorded future path because its question is retrospective: which demonstrator produced this segment? The BC policy instead receives a move_base path and desired velocities, alongside two seconds of lidar, position and inertial history. During dataset construction, the planner prior uses the demonstrated location ten meters ahead. A shared representation feeds two heads: one predicts 200 points of the human path, and the other predicts 20 timesteps of forward and angular commands. Their mean-squared errors are summed for joint training, with no representation pretraining. Thus the predicted future is a robot route, and the learned control is directly supervised by joystick commands. Reader interpretation: sharing these outputs may help representation learning, but no environmental dynamics model or controlled test of that sharing is supplied. e07e09e10e16
5.3 Keep path agreement and physical interaction as separate tests
The paper first asks whether the global head resembles a demonstrated path. Its held-out Hausdorff comparison answers that geometric question, and the crossing-pedestrian image makes one example visible. The physical experiment asks something different: how people experience the learned local planner when it passes a stationary person or approaches a person walking toward it. Fourteen participants rate social compliance and safety after randomized algorithm presentations. Those ratings favor BC over move_base, supporting limited real-world validation beyond an offline path metric. However, the paper does not document the full runtime connection between its global and local heads, nor does it test the complete diversity of the collected scenes. Reader interpretation: these evidence layers are complementary, but cannot be combined into an unreported end-to-end success rate or a general safety guarantee. e10e11e12e13e14e16
5.4 Training and inference
During training
The classifier uses binary cross-entropy on sixteen same-route trajectories from two demonstrators: twelve for training and four for validation. The result paragraph calls the held-out data a test set; no additional test partition is established. e07e08
BC jointly minimizes the sum of the global and local mean-squared-error objectives. Encoders receive no representation-learning pretraining; no frozen module is described. The paper reports providing a 75%/25% trajectory train/test split for Spot, but does not give optimizer, learning-rate, batch-size, training-duration or compute details. e10e15
During inference
Global path prediction is evaluated offline against recorded human paths. Physical validation deploys the learned local planner on Spot against move_base. The paper does not establish how the predicted global head is connected to the deployed controller, how deployment goals replace recorded future locations, or how the 20 predicted commands are selected and refreshed. No learned environmental rollout or inverse-dynamics action extraction is described. e09e10e11e12
5.5 Implementation flow
- Collect natural demonstrations
Four operators, including the first two authors, teleoperate within UT Austin. Each walks behind the robot at an average distance of two meters without explicitly interacting with the crowd. Collection includes buildings, sidewalks, roads, lawns and crowded football game days; the sampled population is primarily campus users. e03
- Record embodiment-specific sensing
Both robots record 10 Hz VLP-16 lidar, 16 Hz inertial sensing and 20 Hz front RGB imagery. Jackal adds 20 Hz stereo and 30 Hz wheel odometry; Spot has five 5 Hz monocular cameras, visual odometry and leg joint angles. Streams and sensor-to-body transforms are recorded using the AMRL stack in rosbag format. e04
- Attach scenario labels
Each trajectory receives a list drawn from twelve interaction labels. These include street crossing, overtaking, narrow doorways and large crowds. Counts are uneven: With Traffic has 74 tags, while Blind Corner and Crossing Stationary Queue each have six. These are trajectory descriptions, not tracked pedestrian identities or dense temporal annotations. e05
- Classify demonstrator style
Ten-second windows combine 1 Hz grayscale bird's-eye-view lidar, relative robot positions, inertial readings, joystick values and a 200-point recorded future path. A four-layer convolutional encoder and a three-layer dense sensor encoder feed a three-layer classifier. This is retrospective identity classification, not a deployable navigation policy. e07
- Learn paths and commands together
The BC model uses a shared observation representation with separate three-layer global and local heads. Two seconds of lidar at 2 Hz, relative positions and inertial information are combined with move_base's path and desired velocities. Training priors use the demonstrated robot location ten meters ahead. Outputs are a 200-point path within that horizon and 20 command timesteps; v denotes forward velocity and ω angular velocity. e09e10
6. Experiments & results
SCAND records human joystick control alongside robot sensing so social navigation can be learned from demonstrated behavior. It contains 138 trajectories, 8.7 hours and approximately 40 km (reported also as 25 miles), collected by four demonstrators using Jackal and Spot. Its preliminary Spot experiments distinguish demonstrators, imitate future paths and velocity commands, and obtain favorable participant ratings against move_base. These establish useful demonstration data and limited policy validation, rather than comprehensive social-navigation performance (e02–e14).
The paper contains no controlled component ablation. Figure 5 fills the diagnostic role with a qualitative path example and is explicitly not treated as an ablation. Figure 4 is the classifier architecture; the related BC architecture is described only in its caption and method text, so no BC diagram is invented. Tables I and II quantify dataset comparison and composition, not policy performance; experimental performance is reported in prose and Figure 7. The physical study covers two simple Spot scenarios and does not establish cross-robot transfer or a comprehensive benchmark. e05e06e07e09e11e12e13e14e16
6.1 Read the original evidence
Table I. SCAND's contribution combines demonstrations and sensing across two robot morphologies; collection size is only one comparison axis. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the bottom row and read across: SCAND contains 138 trajectories, 40 km and 522 minutes, uses teleoperation, and covers indoor and outdoor spaces with two robots. Then compare the sensor and navigation-method columns before comparing duration. Section II explains that SCAND targets socially intended motion demonstrations, whereas the other resources also serve localization or perception. JRDB's 54 trajectories and 64 minutes provide one relevant scale comparison. CoBot and NCLT show why the table cannot support a claim that SCAND is largest on every measure. Treat each row as a source-reported description, not a common benchmark protocol. e02e06
What it supports. The table supports the dataset's scale and breadth of recorded modalities. Its 522 minutes correspond to the 8.7 hours stated in the abstract. The research opportunity comes from pairing observations with human-chosen robot behavior; a longer collection designed for another task is not automatically a stronger social-navigation training set.
Where the evidence stops. These datasets differ in purpose and collection procedure. The surrounding prose broadly describes teleoperation, but the table marks CoBot and FLOBOT autonomous. Preserve that source inconsistency rather than treating the prose as a correction to those rows.
Table II. Trajectory tags reveal which encounters are represented and which are scarce. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the middle column before interpreting the counts. Narrow Doorway means a doorway encounter where the robot waits for a person to open the door; it is more specific than simply passing through a small opening. Passing Conversational Groups describes moving past people talking together. Section III-B says that trajectories carry lists of these labels, so the rows should not be interpreted as mutually exclusive classes partitioning the dataset. Compare common With Traffic and Sidewalk entries with the six-tag Blind Corner and Crossing Stationary Queue entries. This helps identify concrete subsets for evaluation without assuming dense event timestamps or pedestrian tracks are supplied. e05e14
What it supports. The label inventory makes social scenarios searchable and exposes uneven coverage. With Traffic has 74 tags and Sidewalk 57, while two categories have only six each. A strong aggregate imitation result therefore would not establish equally strong behavior across all twelve interaction types.
Where the evidence stops. Tag counts are not independent trial counts or measures of social success. The source literally defines With Traffic as navigating 'with oncoming traffic'; its directional wording is ambiguous. No annotation-agreement study or human detection/tracking labels are supplied in the paper.
Figure 7. Participants preferred the BC local planner in simple indoor encounters; the scores describe their perceptions. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the legend first: blue is move_base and orange is the BC agent. The left two pairs concern social compliance, while the right two concern safety; each measure has separate static and dynamic conditions. In the static setup, a person stands between the robot's start and goal. In the dynamic setup, the person and robot move toward each other's starting positions. Fourteen participants rated each algorithm from one to five, with algorithm order randomized. The black bars show standard deviations according to the caption, so they are not confidence intervals and may extend beyond the questionnaire's upper response value. e12e13
What it supports. Orange means exceed blue in all four displayed comparisons. The text reports pooled social compliance as BC 4.39 versus move_base 2.86, and safety as BC 4.71 versus 2.89. These author-reported summaries are retained. However, the BC social-compliance bars are approximately 4.85 in the static scenario and 4.28 in the dynamic scenario: a simple equal-weight average is around 4.57, rather than 4.39. The paper provides no raw ratings or pooling details to reconcile this difference. These graph estimates diagnose the source discrepancy; they are not corrected official results.
Where the evidence stops. The outcomes measure felt safety and consideration, not collision rates. move_base is not designed for social compliance. The paper reports significant one-way ANOVA results but does not explain within-participant dependence; this matters when interpreting repeated ratings from fourteen people.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Demonstrator identity classification Spot; two demonstrators on Speedway road; sixteen trajectories with a 12/4 training/held-out division; ten-second windows. | 74.48% Held-out classification accuracy | 50% random guessing, as stated by the authors. Supports distinguishable recordings. Interpreting the difference specifically as social strategy assumes nuisance cues such as session conditions do not explain classification; no uncertainty is reported. e07e08 |
| Offline global path imitation Held-out Spot data; predicted path compared with the demonstrated future path; paper supplies a 75%/25% trajectory split. | BC: 0.26 Average Hausdorff distance; lower is better; unit not stated | move_base: 1.25 Measures geometric agreement with demonstrations, not independent safety or closed-loop navigation success. The paper does not specify uncertainty, a detailed metric implementation or per-scenario breakdown. e10e11 |
| Perceived social compliance during physical Spot navigation Fourteen indoor participants; static obstacle-person and dynamic passing scenarios; algorithm order randomized; ratings summarized across scenarios. | BC: mean 4.39, SD 0.99 Participant rating from 1 to 5; mean and standard deviation | move_base: mean 2.86, SD 0.82 Participants rated BC more considerate. The authors report one-way ANOVA p < 0.001. The reported social-compliance mean of 4.39 is retained, but it is not reconstructed by a simple equal-weight average of Figure 7's approximately 4.85 static and 4.28 dynamic BC bars. The paper provides no raw ratings or pooling details to reconcile this discrepancy. The displayed preference for BC remains; the comparator is not designed for social compliance, and this is a small subjective study. e12e13 |
| Perceived safety during physical Spot navigation The same fourteen-participant, two-scenario indoor study; randomized algorithm order. | BC: mean 4.71, SD 0.70 Participant rating from 1 to 5; mean and standard deviation | move_base: mean 2.89, SD 1.18 The authors report one-way ANOVA p < 0.001. Feeling safer does not measure collision probability, minimum clearance or safety under distribution shift. e12e13 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 5. A crossing-pedestrian example shows BC following the demonstrated route more closely than move_base. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the panels in their labeled order. The green rectangle locates the robot; the white annotation in frame 1 identifies the crossing pedestrian. The caption supplies the color key: green is the human demonstration, red is move_base and yellow is the learned BC path. In frame 2, the red curve bends toward the pedestrian's future position while the yellow path stays near the green demonstration. These are competing robot paths viewed against lidar observations, not predicted pedestrian trajectories. Section IV-B separately evaluates the global head by Hausdorff distance to the demonstrated future path on held-out data. e09e10e11e16
What it supports. The example illustrates the behavior behind the offline result: average Hausdorff distance is 0.26 for BC versus 1.25 for move_base. The numeric comparison comes from the accompanying results paragraph, not from measuring this image. It supports closer geometric imitation of demonstrated routes under the reported setup.
Where the evidence stops. This is a qualitative diagnostic, not a controlled ablation. The source provides neither a scale bar here nor a unit for the reported distance. Path similarity and this selected example cannot establish collision avoidance or isolate the benefit of the move_base prior.
7. Analysis & limitations
7.1 What the evidence leaves open
Collection in one city can encode regional conventions, including keeping right and overtaking on the left. Four demonstrators and campus-heavy demographics limit diversity; rare interactions remain difficult. The authors explicitly leave broader benchmarking and more sophisticated social scenarios for future work. e03e14
No controlled ablation isolates temporal history, shared heads, the move_base prior or pretraining. Figure 5 is an illustrative diagnostic. Validation uses Spot, so the presence of Jackal demonstrations does not establish cross-embodiment transfer. e09e10e11e16
The classifier does not isolate social style from all recording cues. Likewise, repeated ratings from the same participants motivate a participant-aware analysis; the reported one-way ANOVA does not explain how within-participant dependence is handled. e07e08e12e13
7.2 Questions for discussion
- Would demonstrator identity remain predictable after controlling recording-session and crowd differences (e07–e08)?
- How should a deployment goal replace the recorded ten-meter future position used to construct training priors (e09–e10)?
- Which rare social tags and local conventions most challenge generalization beyond Austin (e05, e14)?
8. Reproducibility audit
8.1 Requirements and known gaps
Rebuilding the baselines requires rosbag decoding, frame transforms, lidar-to-BEV preprocessing, aligned inertial/odometry/action windows, trajectory-level splits and move_base priors. The paper links a public dataset and software but states no dataset license; exact preprocessing, software versions and controller settings are not fully specified. e02e04e07e09e15
Reader-proposed checks: hold routes and recording sessions apart when testing demonstrator classification, and compare BC with and without move_base inputs under identical splits and training budgets. Report trajectory-level uncertainty and keep offline imitation metrics distinct from any subsequent physical evaluation. e07e08e09e10e11
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Test whether demonstrator classification survives session controls
Reader-proposed, not performed: first reconstruct the sixteen-trajectory same-route classifier experiment and its 12/4 split, ensuring all windows from one trajectory stay together. Then evaluate session-disjoint folds, matching route portions and balancing demonstrator labels as far as the available metadata permits. Keep architecture, window length, preprocessing and training budget fixed; compare with the stated 50% guessing baseline and a label-permutation control. Report per-trajectory accuracy and uncertainty across held-out trajectories. Accuracy remaining above chance across sessions would support a stable behavioral signature; a collapse limited to session-disjoint testing would weaken the social-style interpretation. Missing session identifiers or insufficient matched data must be reported as a feasibility limitation. e07e08e15
Check 2: Isolate the move_base prior before attributing BC's path improvement
Reader-proposed, not performed: on the same Spot 75%/25% trajectory split, compare full BC against BC trained from scratch with only the move_base path and desired-velocity inputs removed. Keep histories, output heads, loss definitions, training budget and random seeds matched. Include raw move_base as the original baseline. Measure held-out Hausdorff distance using an explicitly documented unit and implementation, and report trajectory-level uncertainty and tag-stratified results where counts permit. Improvement by full BC over the no-prior version would support the authors' prior hypothesis; indistinguishable results would weaken it. Document the demonstrated ten-meter goal used for offline priors, and separately test deployable goal selection before drawing online-control conclusions. e05e09e10e11e15e16
8.3 Reading coverage
Visual audit: All eight PDF pages were rendered and visually inspected, covering the title and revision block, collection and sensor specifications, classifier and BC descriptions, evaluation protocols, results, limitations and references. Figures 1–7 and Tables I–II were read. All six final original crops were separately inspected; table headers, sensor labels, diagram arrows, path annotations, rating axes and legends are retained. Figure 4 was cross-checked against its caption and Sections IV-A/IV-B: it depicts only the classifier. Figure 5's colors were checked against its caption; Figure 7's error bars were checked against its caption and the p. 6 protocol. The embedded raster detail in Figures 4, 5 and 7 remains limited by the source; higher-DPI PDF rendering does not create additional measurements. No appendix is present. The supplementary video, dataset files, linked code and external references remain outside this review.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8. Appendix coverage: not present.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- PDF p. 1: title, author affiliations, revision labels and abstract
- PDF pp. 1–2: I. Introduction
- PDF pp. 2–3: II. Related Work, including learning methods and dataset comparisons
- PDF pp. 3–4: III. Data Collection Procedure, A. Collecting Data and B. Labeled Annotations of Social Interactions
- PDF pp. 4–6: IV. Analysis, A. Demonstrator Classification and B. Imitation Learning for Global and Local Planning
- PDF p. 7: V. Anticipated Use Cases and VI. Conclusion
- PDF pp. 7–8: Acknowledgments and References
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- The extraction limitation was addressed by visually inspecting all eight PDF pages and every final crop. The supplementary video, dataset files, linked software and external references were not inspected; no experiments were reproduced.
- Edition identity: the supplied PDF is arXiv:2203.15041v2, dated 8 June 2022, labeled an IEEE Robotics and Automation Letters preprint. Its header says accepted June 2022; its manuscript footnote says accepted 22 May 2022. Both labels are retained without reconciliation. The catalog's journal pagination and final publisher edition were not verified.
- The observed title differs from the catalog only in capitalization. All nine authors agree in identity and order; the title page credits Justin Hart, whereas the catalog uses Justin W. Hart. Metadata preserves the observed title-page form. No earlier revision or publisher edition was compared.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block, author affiliations, arXiv margin and manuscript footnote
Observed title and nine authors match the catalog work; title-page name is Justin Hart. The margin identifies arXiv:2203.15041v2, 8 June 2022. Header says preprint accepted June 2022; footnote says accepted 22 May 2022. Affiliations are UT Austin Mechanical Engineering, UT Austin Computer Science, Robotics@Google, Sony AI and Army Research Laboratory's Computational and Information Sciences Directorate.
Go to primary source ↓e02PDF p. 1, Abstract and Section I; p. 2, Section I and dataset footnote
SCAND supplies 8.7 hours, 138 trajectories and 25 miles of human-teleoperated demonstrations from four operators on Jackal and Spot, with public-release provenance. Motivation is learning social navigation from observations and commands in natural environments.
Go to primary source ↓e03PDF p. 3, Section III-A; p. 4, Section III-A first paragraph
Four demonstrators collect data around UT Austin, walking on average two meters behind the robot. Crowds are mainly campus denizens; routes include indoor and outdoor spaces and two football game days. Scenarios are not orchestrated.
Go to primary source ↓e04PDF p. 4, Section III-A sensor and storage paragraphs; p. 5, Figure 3 and caption
Common sensors are 10 Hz VLP-16 lidar, 16 Hz IMU and 20 Hz front RGB. Jackal adds 20 Hz stereo and 30 Hz wheel odometry; Spot has five 5 Hz monocular cameras, visual odometry and joint angles. Frame transforms and rosbag recording through the AMRL stack are described.
Go to primary source ↓e05PDF p. 4, Section III-A final paragraph, Section III-B, Table II all rows and Figure 2
Twelve predefined tags describe interactions along trajectories. Counts are Against Traffic 22, With Traffic 74, Street Crossing 34, Overtaking 14, Sidewalk 57, Passing Conversational Groups 38, Blind Corner 6, Narrow Doorway 15, Crossing Stationary Queue 6, Stairs 22, Vehicle Interaction 21 and Navigating Through Large Crowds 27. No human detection/tracking annotations are provided. Table II describes With Traffic as navigating with oncoming traffic.
Go to primary source ↓e06PDF p. 3, Table I, SCAND, JRDB, CoBot, NCLT and FLOBOT rows; Section II-B.2
Table I lists SCAND as 138 trajectories, 40 km, 522 minutes, teleoperated, two robots and indoor/outdoor coverage; JRDB has 54 trajectories and 64 minutes. Other datasets exceed SCAND on some size measures. CoBot and FLOBOT are labeled autonomous in the table, despite the surrounding prose's broad statement about teleoperation.
Go to primary source ↓e07PDF p. 5, Section IV-A.1 and Figure 4 with caption
Classifier uses sixteen same-route trajectories from two demonstrators, twelve training and four validation. Ten-second windows include lidar at 1 Hz, relative positions, a 200-point future path, IMU and joystick values. Four-layer CNN and three-layer dense encoders feed a three-layer dense classifier trained with binary cross-entropy. Figure 4's arrows run from the two encoders into the classifier and demonstrator label.
Go to primary source ↓e08PDF p. 5, Section IV-A.2
Held-out demonstrator accuracy is 74.48%, compared with 50% random guessing. Authors interpret this as multiple socially compliant strategies; the result paragraph calls the held-out data a test set.
Go to primary source ↓e09PDF p. 5, Figure 4 caption and Section IV-B.1; p. 6, Section IV-B.1 continuation
The BC model is not drawn in Figure 4. It uses a similar shared observation representation with separate three-layer global/local heads, two-second observation windows and 2 Hz BEV lidar. Inputs include relative positions, inertial information and move_base global path/desired velocities, constructed using the robot's demonstrated ten-meter future location.
Go to primary source ↓e10PDF p. 6, Section IV-B.1 first paragraph and Section IV-B.2 final paragraphs
Global head predicts 200 path points; local head predicts 20 timesteps of joystick commands (v, ω). Their MSE losses are summed and parameters updated together without encoder representation pretraining. A 75%/25% trajectory split for Spot is reported. Command execution cadence and global/local deployment coupling are not specified.
Go to primary source ↓e11PDF p. 6, Section IV-B.2 first paragraph and Figure 5 with caption
Average held-out Hausdorff distance is 1.25 for move_base and 0.26 for BC, without a stated unit. Figure 5 shows three frames: green is the human-demonstrated path, red move_base and yellow BC; in frame 2 the red path turns toward the pedestrian's future state while BC resembles the demonstration.
Go to primary source ↓e12PDF p. 6, Section IV-B.2 physical-study protocol, Figure 6 and questionnaire
Fourteen participants evaluate Spot indoors. Static trials place a stationary human five meters from both start and goal; dynamic trials have human and robot exchange positions initially ten meters apart. Algorithm order is randomized. Participants rate social compliance and perceived safety from 1 to 5.
Go to primary source ↓e13PDF p. 6, Section IV-B.2 paragraph beginning 'We randomized'; p. 7, Figure 7 and caption
Reported pooled social-compliance means/SDs are BC 4.39/0.99 and move_base 2.86/0.82; safety means/SDs are BC 4.71/0.70 and move_base 2.89/1.18. Authors report one-way ANOVA: safety F(1,55)=47.87 and social compliance F(1,55)=38.67, both p<0.001. Figure 7 separates static/dynamic bars and displays mean and standard deviation, not confidence intervals.
Go to primary source ↓e14PDF p. 6, final paragraph of Section IV-B.2; p. 7, Section V
Authors limit demonstrated BC ability to simple scenarios, identify rare interactions and single-city regional bias, and place broader benchmarking, representation learning, real-to-sim transfer and inverse-reinforcement-learning uses in future work.
Go to primary source ↓e15PDF p. 2, release footnote; p. 4, Section III-A; pp. 5–6, Sections IV-A.1 and IV-B.1–2
The source describes storage, sensor rates, network depths, objectives and trajectory splits, but does not specify a dataset license, full BEV rasterization/alignment procedure, optimizer, learning rate, batch size, compute budget, software versions or complete execution configuration.
Go to primary source ↓e16PDF p. 4, Section IV opening; pp. 5–6, Sections IV-A and IV-B; p. 7, Section V
Analysis is explicitly based on Spot data and consists of demonstrator classification and BC path/local-policy validation. There is no controlled component ablation or cross-robot transfer experiment; prior inputs and representation pretraining are discussed as hypotheses or future improvements.
Go to primary source ↓8.5 Primary sources
Socially CompliAnt Navigation Dataset (SCAND): A Large-Scale Dataset Of Demonstrations For Social Navigation ↗
PDF · 6,973 extracted words
Source fingerprint
b82d5775bc1f2f0ab93614a3e55d7ade6b3a3256a5d74ae343fad326c2ba720d