CARLA: An Open Urban Driving Simulator
1. Paper overview
In one sentence: CARLA makes urban-driving experiments controllable and diagnosable, while its original baselines expose a large gap between reaching destinations in familiar scenery and driving reliably in a new town. e-identitye-platforme-protocole-completione-infractions
| At a glance | What to know |
|---|---|
| Research problem | Source description Physical urban-driving research is expensive, and dangerous corner cases are difficult to stage. CARLA offers configurable traffic, sensors and weather so researchers can train and evaluate perception and closed-loop control in controlled simulated scenarios. e-identitye-platform |
| Core mechanism | Source description An Unreal Engine 4 client–server simulator combines purpose-built urban assets with sensor readings, control commands and detailed infraction feedback. The authors describe its code, protocols and assets as open; the PDF does not establish a currently installable release. e-platforme-apie-conclusion |
| A key reported result | Navigation with dynamic obstacles — town generalization: Town 1: MP 77%, IL 83%, RL 7%. Town 2: MP 24%, IL 38%, RL 2%. Successfully completed episodes (%); higher is better.. Same trained agents; cars and pedestrians present. Compare Town 1 and unseen Town 2 under the Training Weather Set. Section 4 states 25 episodes per task/town/weather-set combination. IL exceeds MP by 6 percentage points in Town 1 and 14 in Town 2; RL has the lowest completion. Town transfer substantially degrades every tested implementation. Success means reaching the destination within the timeout, even with infractions; no uncertainty estimates are reported. e-protocole-completion |
| Reading caution | Reader analysis Only two towns and six evaluation weathers are tested. Town transfer changes layout, textures and models together; routes also differ in length. These experiments neither isolate the cause of transfer failure nor establish physical-road performance. e-townse-protocole-discussion |
Core contributions
- Source description
An Unreal Engine 4 client–server simulator combines purpose-built urban assets with sensor readings, control commands and detailed infraction feedback. The authors describe its code, protocols and assets as open; the PDF does not establish a currently installable release. e-platforme-apie-conclusion
- Source description
A benchmark separates route complexity, dynamic obstacles, town transfer and weather transfer, comparing three driving approaches using both destination completion and distance between infractions. e-protocole-completione-infractions
Figure 1. Weather controls alter the visual conditions of the same urban setting. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the panels clockwise from the upper left: clear day, daytime rain, daytime after rain, and clear sunset. The buildings and road identify a common setting, while rainfall, reflections, shadows and illumination change what a camera sees. This illustrates a controllable environmental factor in CARLA, rather than the output of a learned video predictor. The client can select environmental settings separately from vehicle controls. In the benchmark, four weather conditions are used during training and two others are withheld, creating a weather-transfer test inside either town. The figure illustrates the rendering capability; the numerical test is in Table 1. e-weathere-apie-protocole-completion
What it supports. Weather variation is an explicit experimental control. Table 1 shows that MP and IL can perform well under the selected unseen weathers in Town 1, whereas changing town creates a larger difficulty. The pictures help explain what the weather factor changes, but do not by themselves establish policy robustness.
Where the evidence stops. These are illustrative simulator views, not a quantitative realism test. The paper reports 18 illumination–weather combinations in Section 2, but S.1.1 lists 14 presets; the correspondence is unresolved. The six evaluation weathers are specified separately.
2. Motivation
2.1 The problem and the proposed response
Physical urban-driving research is expensive, and dangerous corner cases are difficult to stage. CARLA offers configurable traffic, sensors and weather so researchers can train and evaluate perception and closed-loop control in controlled simulated scenarios. e-identitye-platform
2.2 What this reading follows
A driving simulator becomes scientifically useful when researchers can control what changes and measure how a policy fails. CARLA connects configurable urban worlds to a client that receives sensor observations and sends vehicle controls. The original paper tests a modular pipeline, conditional imitation learning and reinforcement learning across route difficulty, traffic, weather and town changes. Read its two result tables together: reaching a destination does not require avoiding collisions in this protocol. The illustrations below connect those measurements to the available sensors, the imitation policy's architecture and the environmental split. This edition covers the supplied CoRL 2017 paper and its supplement. e-identitye-platforme-protocole-completione-infractions
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 | Benchmarks & simulators |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded benchmark/simulator category is supported: CARLA supplies a simulated world, interfaces and evaluation protocols. Its three baseline policies do not make CARLA a joint future/action predictor. One Model, prediction-paradigm and WAM-quadrant labels are therefore not applicable, rather than merely unverified. e-platforme-routee-protocol
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 |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Separate the simulated world from the driving policy
CARLA's server owns the urban scene and advances the simulation; the client supplies controls and receives observations. This separation is the starting point for interpreting the paper. Camera images are observations of an executed simulated state. Depth and segmentation pseudo-sensors expose privileged scene information, and state measurements provide feedback for training or evaluation. The A* planner adds route intent, but it does not solve the local driving problem: its commands lack geometric trajectories. A policy must still convert observations and the intended intersection maneuver into steering, throttle and brake. Reader interpretation: CARLA is infrastructure for studying world interaction, while the baseline policies are the agents under study. Calling the simulator a learned world-action model would erase this architectural distinction. e-platforme-sensorse-apie-route
Figure 2. CARLA separates rendered appearance from ground-truth scene information. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read left to right: an RGB camera view, ground-truth depth, then ground-truth semantic segmentation. Match the road, vehicles and pedestrian across panels to see how the same scene is exposed through different observation channels. The depth and segmentation panels are pseudo-sensor outputs supplied by the simulator; they are not predictions made by the driving policy. Section 2 also describes state and infraction measurements outside these images. This distinction matters for experiments: a learned perception module can be evaluated against privileged information, and a downstream planner can potentially be tested with perception errors removed. Such an intervention is proposed below, not reported in the paper. e-sensorse-mpe-route
What it supports. The interface provides both perceptual challenges and information that can help diagnose them. MP normally estimates segmentation from images; Figure 2 shows why CARLA can support a separate ground-truth-perception control. Availability of a pseudo-sensor does not mean every baseline receives its output during driving.
Where the evidence stops. The colored depth image has no numerical color scale, so no depth values are inferred. The panel is not evidence of segmentation accuracy. Baseline input descriptions must be read separately from the simulator's full list of available signals.
5.2 Track where each baseline obtains its competence
The modular pipeline learns perception, then relies on hand-designed local planning and PID control; its segmentation back end is pretrained and frozen. IL learns from demonstrations and selects a motor-prediction branch using the route command. Its demonstrations are not exclusively human: the supplement gives an 80% automated and 20% human mixture, with the automated expert receiving privileged state. A3C instead learns from rewards and its own interaction. Its reward equation uses changes in speed, collision damage and road overlap, as well as a goal-distance term. Reader interpretation: the comparison combines differences in supervision, inputs, regularization and engineering. The paper's weaker RL results cannot isolate which of these factors dominates, even though all three implementations face the same families of driving tasks. e-mpe-il-architecturee-il-traininge-rle-rewarde-protocole-discussion
Table S.1. The imitation policy fuses image and speed features before command-dependent control. Original paper, p. 15 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the Perception group, whose first row specifies the 200×88×3 image input. The lower rows produce a 512-dimensional visual representation. Read Measurement separately: its scalar speed input becomes a 128-dimensional feature vector. The Joint input row explicitly combines 512 and 128 dimensions and produces 512 features for control. Section S.2.2 supplies the missing routing explanation: four identical command-conditional branches are available, and one is selected by the navigation command. The table does not depict this selector. Its printed dimensions are retained exactly, while the prose establishes the verified high-level flow from image and speed to actuator prediction. e-il-architecturee-route
What it supports. IL uses the route command to choose a control branch, allowing the same visual intersection to support different intended maneuvers. The image and speed encoders feed motor prediction directly. Neither this table nor the accompanying architecture description specifies a learned future-world rollout for planning.
Where the evidence stops. The final Control row lists one channel, whereas S.2.2 describes steering, brake and throttle. The source does not reconcile that actuator-head layout. Layer dimensions remain as printed; this crop should not be treated as a fully validated executable network specification.
5.3 Ask what a successful episode actually establishes
The benchmark first asks whether an agent reaches its goal within a route-dependent time budget. It separately records infractions, allowing an agent to finish despite collisions or sidewalk driving. Consequently, Table 1 and Table 2 answer complementary questions. In familiar conditions, IL has higher completion on dynamic navigation, while MP has a longer distance between car collisions. A town change reduces the strongest completion rates sharply, but the weather change is not uniformly harder. Reader interpretation: the defensible conclusion is a profile of strengths under the specified protocol, not a single universal ranking of driving methods. The absence of reported uncertainty and the unexplained percentage aggregation also discourage strong claims about close numerical differences. Collision-free scoring would be a new analysis, proposed below. e-protocole-completione-infractions
5.4 Training and inference
During training
MP segmentation uses 2,500 labeled simulator images, Adam for 300 epochs at learning rate 10^-6 and batch size 2, with an ImageNet-pretrained frozen ResNet back end. The intersection classifier uses 500 balanced images, 500 epochs, learning rate 10^-3 and batch size 5, without pretraining. Neither uses augmentation. e-mp
IL uses about 14 hours of demonstrations. Section 3.2 describes human driving, but S.2.2 specifies 80% privileged automated-agent demonstrations and 20% human demonstrations. Steering perturbations enrich collection. Training uses command-balanced batches of 120 for 294,000 iterations, Adam starting at 0.0002 and halved every 50,000 iterations, plus image augmentation and dropout. e-il-training
RL uses 10 actor threads and 10 million environment steps, with 20-step rollouts, initial learning rate 0.0007 decaying linearly to zero, and entropy regularization 0.01. Collisions terminate RL training episodes. Its lack of IL-style augmentation/dropout and different supervision make this an implementation comparison rather than an isolated learning-paradigm experiment. e-rle-discussion
During inference
Each policy repeatedly converts current observations and route intent into controls; the simulator executes them and supplies feedback. IL selects a branch by command; MP updates waypoints and PID actions. None of these baselines uses learned future-world prediction for inference-time planning. e-routee-mpe-il-architecturee-rl
Evaluation reuses one trained agent per approach across all tasks without task-specific fine-tuning. Unlike RL training, evaluation continues after collisions and other infractions. Agents may ignore traffic lights and speed limits. e-protocole-rl
5.5 Implementation flow
- Run the environment
The UE4 server advances and renders the world; a Python client exchanges controls, resets and sensor data over sockets. The paper supplies two towns: Town 1 has 2.9 km of drivable roads and Town 2 has 1.4 km. RGB cameras and ground-truth depth/segmentation pseudo-sensors are configurable. e-platforme-townse-sensors
- Provide route intent
All agents receive high-level commands from an A* topological planner using current and goal locations. It specifies intersection choices without a geometric trajectory or metric map; each policy must still perceive and execute local motion. e-route
- Modular pipeline (MP)
RefineNet-based segmentation estimates road, lane markings and hazards; an AlexNet-based classifier detects intersections. A five-state rule-based planner produces waypoints, and PID control targets 20 km/h. Left turns additionally use an auxiliary camera. This is an engineered perception–planning–control chain. e-mp
- Conditional imitation learning (IL)
A 200×88 image produces a 512-dimensional representation; speed produces a 128-dimensional representation. Their fusion feeds four command-selected control branches. The prose specifies steering, brake and throttle, while Table S.1 ends in a scalar channel; the precise actuator-head arrangement remains ambiguous. e-il-architecture
- Reinforcement learning (RL)
A3C processes the two most recent 84×84 images separately from speed, distance to goal, collision damage and a one-hot route command, then combines the representations. It learns action selection from simulated reward, without demonstrations. e-rl
6. Experiments & results
CARLA supplies controllable urban simulation and evaluation signals for driving policies. Its original benchmark compares modular control, conditional imitation learning and A3C, exposing large failures on a new town even when unseen weather is manageable. Completion and infraction metrics reveal different strengths.
The source contains no simulator/policy block diagram and no controlled component-ablation table. Figure 2 explains the observation interface, and Table S.1 supplies the original architecture specification. Table 2 is included in the ablation section as a behavioral diagnostic; it does not isolate a mechanism. The paper's environment images and comparative tables support six original visuals without inventing an ablation. e-weathere-sensorse-townse-il-architecturee-completione-infractionse-discussion
6.1 Read the original evidence
Figure S.1. The held-out town changes both road structure and visual surroundings. Original paper, p. 13 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the entire left column as Town 1 and the entire right column as Town 2. The top panels show road layouts, while the lower panels illustrate the streets encountered by a camera. Training uses Town 1; Town 2 is held out. These are maps for understanding the evaluation environments, not metric maps supplied to the baseline policies. The policies instead receive high-level intersection commands. Compare this figure with the column groups in Table 1: moving to New town changes the environment while retaining training weather, whereas New weather retains Town 1 and changes the weather set. e-townse-routee-protocole-completione-discussion
What it supports. On navigation with dynamic obstacles under training weather, IL drops from 83% completion in Town 1 to 38% in Town 2, while MP drops from 77% to 24%. The visual split makes clear why generalization across towns is more than seeing a familiar street under another sky.
Where the evidence stops. The town comparison changes layout and appearance together, and average route lengths differ across towns. The result cannot isolate which change causes failure. The maps and views also provide no evidence of transfer to physical streets.
Table 1. Destination completion reveals a strong dependence on environment transfer. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Choose a task row first, then compare MP, IL and RL inside one condition group. Next hold the method fixed and move across groups to examine transfer. MP denotes the modular pipeline, IL conditional imitation learning, and RL the A3C implementation. Values are percentages of episodes that reach the goal before the timeout; larger values are better. Nav. dynamic is navigation with cars and pedestrians. Training conditions still use start and goal locations different from training. The timeout is the optimal-path travel time at 10 km/h. Traffic-light and speed-limit compliance is not required, and collisions do not end evaluation episodes. e-protocole-completion
What it supports. In the hardest row, training-condition completion is MP 77%, IL 83% and RL 7%; in New town it is 24%, 38% and 2%. MP and IL remain substantially stronger than this RL baseline, but neither achieves reliable transfer. New town & weather gives 44%, 42% and 4%, so the two shifts are not monotonically additive.
Where the evidence stops. The text states 25 episodes per task/town/weather-set combination, but the percentages are not all multiples of four. Aggregation and repeated-run details are unclear, and uncertainty is absent. These are destination-completion results, not collision-free success rates or statistical significance tests.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Navigation with dynamic obstacles — town generalization Same trained agents; cars and pedestrians present. Compare Town 1 and unseen Town 2 under the Training Weather Set. Section 4 states 25 episodes per task/town/weather-set combination. | Town 1: MP 77%, IL 83%, RL 7%. Town 2: MP 24%, IL 38%, RL 2%. Successfully completed episodes (%); higher is better. | IL exceeds MP by 6 percentage points in Town 1 and 14 in Town 2; RL has the lowest completion. Town transfer substantially degrades every tested implementation. Success means reaching the destination within the timeout, even with infractions; no uncertainty estimates are reported. e-protocole-completion |
| Navigation with dynamic obstacles — weather generalization Unseen cloudy daytime and soft rain at sunset; Town 1 ('New weather') versus Town 2 ('New town & weather'). | Town 1: MP 89%, IL 82%, RL 2%. Town 2: MP 44%, IL 42%, RL 4%. Successfully completed episodes (%); higher is better. | Under training weather, corresponding values are MP/IL/RL 77/83/7 in Town 1 and 24/38/2 in Town 2. Unseen weather is comparatively manageable for MP and IL, but weather sets differ in difficulty: adding the weather shift does not monotonically lower performance. e-protocole-completion |
| Navigation with dynamic obstacles — infraction diagnosis Training conditions, with infraction definitions in S.3.1; distances accumulated during navigation with dynamic objects. | Car collisions: MP 16.4 km, IL 3.26 km, RL 0.58 km. Opposite-lane violations: MP 10.2 km, IL 33.4 km, RL 0.18 km. Average kilometers traveled between infractions; higher is better. | MP avoids car collisions more frequently per distance, while IL best maintains the correct lane. Completion ranking does not summarize behavior. In new-town pedestrian collisions RL reaches 2.55 km versus MP 1.40 and IL 1.88, but in training conditions MP's 18.9 exceeds RL's 17.8; the discussion's blanket RL pedestrian advantage needs this qualification. e-infractionse-infraction-definition |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 2. Infraction distances reveal behavioral differences hidden by completion rates. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Keep the condition group fixed and compare methods within an infraction row. The unit is kilometers traveled between infractions, so larger values indicate fewer events per distance. These measurements concern navigation with dynamic objects. Opposite-lane and sidewalk events require more than 30% footprint overlap; an ongoing violation is counted again every two seconds. Collision rows refer to contact with static objects, cars or pedestrians. Read the New weather IL sidewalk cell literally as greater than 57, preserving its bound. Then compare with Table 1: the method reaching more destinations can still have more frequent collisions along its trajectories. e-infractionse-infraction-definitione-protocole-completion
What it supports. In training conditions, MP travels 16.4 km between car collisions versus IL's 3.26 and RL's 0.58, while IL leads the opposite-lane metric at 33.4 km versus 10.2 and 0.18. This separates collision avoidance from lane keeping. RL's new-town pedestrian value is strongest, but MP leads that row in training conditions.
Where the evidence stops. The prose broadly says RL collides with pedestrians least often; Table 2 qualifies this, since training-condition MP reaches 18.9 km versus RL's 17.8. This is a behavioral diagnostic, not a controlled ablation. Exposure to pedestrians and uncertainty are not separately reported.
7. Analysis & limitations
7.1 What the evidence leaves open
Only two towns and six evaluation weathers are tested. Town transfer changes layout, textures and models together; routes also differ in length. These experiments neither isolate the cause of transfer failure nor establish physical-road performance. e-townse-protocole-discussion
Table 1 percentages are not all multiples of four despite the stated 25-episode protocol. Weather aggregation, repetitions, raw counts and uncertainty are not sufficiently specified to reconstruct those percentages uniquely. Do not infer significance from small gaps. e-protocole-completion
The main text reports 18 illumination–weather combinations, while the supplementary API list enumerates 14 presets. Together with the demonstration-provenance and architecture-table discrepancies, this prevents treating the PDF as an unambiguous release specification. e-weathere-il-traininge-il-architecture
7.2 Questions for discussion
- How much new-town failure would remain with ground-truth segmentation but unchanged local planning?
- Would rankings change if success also required a collision-free route?
8. Reproducibility audit
8.1 Requirements and known gaps
Preserve the original town/weather split, route definitions, sensor configuration, planner and timeout: optimal-path travel time at 10 km/h. Log completion separately from infractions. Wrong-lane/sidewalk events require more than 30% footprint overlap; a continuing violation is counted again every two seconds. e-protocole-infraction-definition
The PDF gives training schedules but omits a pinned simulator build, hardware configuration, exact evaluation routes/seeds and numerical PID/hazard thresholds. Resolve actuator-head dimensions, reward-distance convention and evaluation aggregation before calling a rebuild exact. External release links alone do not fill these gaps. e-platforme-apie-mpe-il-architecturee-rewarde-protocol
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: Replace only MP segmentation with simulator ground truth
Reader-proposed check, not performed: evaluate the same MP policy in both towns and weather sets with learned segmentation versus CARLA ground truth mapped to MP's five categories. Keep the intersection classifier, auxiliary-camera setup, state machine, PID settings, routes, timeout and actor seeds fixed. Report completion and all five infraction distances with paired episodes and repeated seeds. The authors' perception-failure explanation predicts a larger benefit in the failing new-town conditions. If segmentation quality improves but driving barely changes, that would challenge segmentation as the principal bottleneck and motivate a separate test of intersection recognition or planning. This intervention diagnoses MP; it is not a fair privileged-input competitor for the original baseline ranking. e-sensorse-mpe-apie-protocole-discussione-infraction-definition
Check 2: Rescore identical trajectories for collision-free completion
Reader-proposed check, not performed: rerun the original task/town/weather matrix and retain per-episode goals, timeout, trajectory distance and timestamped collisions. Score each identical trajectory twice: original destination completion, and destination completion with zero collisions of any type. Keep episode termination unchanged during collection so scoring is the only intervention. Publish raw counts, the exact aggregation across weather, and uncertainty over repeated seeds. Compare MP and IL on dynamic navigation: Table 2 motivates testing whether MP's lower car-collision frequency compensates for its lower original completion. A rank reversal would establish protocol sensitivity; an unchanged ranking would refute that specific expectation. The new score supplements, rather than reproduces, the paper's metric. e-protocole-completione-infractionse-infraction-definition
8.3 Reading coverage
Visual audit: Directly inspected the title/author page; all method, training, evaluation and result pages; all supplementary technical pages; Figures 1, 2, S.1 and S.2; and Tables 1, 2 and S.1. Six final original crops were individually viewed. Page 14's asset illustration is present and was inspected although not cropped. Pages 9–10 contain acknowledgements/references and were read as text; their cited external works were not opened. Table S.1 was checked against S.2.2, and Tables 1–2 against the protocol, discussion and infraction definitions. The reward equation was visually checked on page 15. No code, external videos or experiments were inspected or executed.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 11, 12, 13, 14, 15, 16. Appendix coverage: reviewed.
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
- Title, author block, abstract and CoRL 2017 footer (PDF p. 1)
- 1 Introduction
- 2 Simulation Engine
- 3 Autonomous Driving; 3.1 Modular pipeline; 3.2 Imitation learning; 3.3 Reinforcement learning
- 4 Experiments
- 5 Results
- 6 Conclusion
- Acknowledgements and References (PDF pp. 9–10)
- S.1 Simulator Technical Details; S.1.1 Client and Server Information Exchange; S.1.2 Environment
- S.2 Driving Systems Technical Details; S.2.1 Modular Pipeline; S.2.2 Imitation Learning; S.2.3 Reinforcement Learning
- S.3 Experimental Setup; S.3.1 Types of Infractions
Outside the original text pass
- Text extraction does not reconstruct figure pixels or preserve equation/table layout; downstream reading must inspect the retained original PDF.
- The sample images on PDF page 14 require visual reading.
- Availability of external supplements, code, datasets and videos beyond the retained PDF has not been fully verified.
- The first two acquisition limitations were addressed by direct PDF-page and crop inspection, including Figure S.2 on page 14. All five supplied text chunks were read individually; the complete 16-page paper includes the supplementary material.
- Identity/version: the observed title matches exactly. The title page prints German Ros and Antonio López, whereas the catalog uses Germán Ros and Antonio M. López. Author continuity is verified. This report concerns the supplied CoRL 2017 official proceedings PDF identified by its raw hash, not a later CARLA release; no separate revision number is printed.
- Code, linked papers, external videos and external supplements were not inspected; no installation or experiments were performed.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title/author block, abstract, Introduction and conference footer
The observed title is CARLA: An Open Urban Driving Simulator. Authors are Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio López and Vladlen Koltun; affiliations are Intel Labs, Toyota Research Institute and Computer Vision Center, Barcelona. The footer identifies CoRL 2017.
Go to primary source ↓e-platformPDF p. 2, Introduction and Section 2; p. 3, first two paragraphs of Section 2
CARLA uses UE4; its Python client exchanges controls and meta-commands with a simulation/rendering server. The text motivates controlled driving research and describes purpose-built assets.
Go to primary source ↓e-weatherPDF p. 2, Figure 1 and caption; p. 3, Environment, final paragraph; p. 11, S.1.1, Weather Id
Figure 1 shows clear day, rain, after-rain daytime and sunset. The main text states 18 combinations; the supplementary list names 14 weather/lighting presets.
Go to primary source ↓e-townsPDF p. 3, Environment; p. 12, S.1.2; p. 13, Figure S.1; p. 14, Figure S.2
Town 1 has 2.9 km of roads and Town 2 has 1.4 km; maps and views show both environments. Figure S.2 displays car and pedestrian assets.
Go to primary source ↓e-sensorsPDF p. 3, Sensors; p. 4, Figure 2/caption and Section 2 continuation
RGB, ground-truth depth and 12-class semantic segmentation are available. State, collision, overlap and dynamic-object information supports training and evaluation; pseudo-sensors allow perception controls.
Go to primary source ↓e-apiPDF p. 11, S.1.1, Commands, Meta-commands, Measurements and sensor readings
The interface exposes steering, throttle, brake, hand brake and reverse gear, plus actor counts, weather, spawn seeds and cameras; feedback includes position, speed, collisions and overlap.
Go to primary source ↓e-routePDF p. 4, Section 3, observation/action notation and topological-planner paragraph
Actions contain steering, throttle and brake. A* supplies high-level intersection choices from current and goal locations; it provides neither geometric trajectories nor metric maps.
Go to primary source ↓e-mpPDF p. 5, Section 3.1, Perception, Local planner and Continuous controller; p. 12, S.2.1
RefineNet and AlexNet modules use 2,500 and 500 images; five planner states feed PID control targeting 20 km/h. Supplement supplies frozen ResNet, optimizer schedules, auxiliary left-turn camera and an unspecified hazard threshold.
Go to primary source ↓e-il-architecturePDF p. 12, S.2.2 Architecture; p. 15, Table S.1
Image and speed branches produce 512- and 128-dimensional representations. Four command-dependent branches predict controls according to the prose; the table lists a final scalar channel and preserves its printed layer dimensions.
Go to primary source ↓e-il-trainingPDF p. 5, Section 3.2; p. 12, S.2.2 Training details; p. 15, continuation and Training data
Main text reports about 14 hours and describes human traces; supplement specifies 80% automated and 20% human demonstrations, privileged expert state and steering perturbations. It gives batch size, Adam schedule, dropout, image augmentation and cropping.
Go to primary source ↓e-rlPDF pp. 5–6, Section 3.3; p. 15, S.2.3
A3C terminates training on goal, collision or timeout. It uses two 84×84 frames and speed, distance-to-goal, damage and route-command measurements, with 10 threads, 10 million steps, 20-step rollouts, learning rate 0.0007 and entropy coefficient 0.01.
Go to primary source ↓e-rewardPDF p. 15, S.2.3, input paragraph, reward paragraph and Equation (2)
Equation (2) uses goal-distance decrease and changes in speed, damage and footprint overlaps. Reward prose describes d as distance traveled towards the goal; the input paragraph names distance to goal.
Go to primary source ↓e-protocolPDF p. 6, Section 4 and opening of Section 5
Four tasks cross two towns and two weather sets (four training and two test weathers). One agent per approach is reused. The text specifies 25 episodes per task/town/weather-set combination, timeout from optimal travel at 10 km/h, no termination for infractions, and permission to ignore lights/speed limits. Training-condition evaluation uses different start/goal positions.
Go to primary source ↓e-completionPDF p. 7, Table 1, all column groups and Nav. dynamic row; Section 5
Nav. dynamic MP/IL/RL percentages are 77/83/7 under training conditions, 24/38/2 in new town, 89/82/2 in new weather, and 44/42/4 in new town and weather. The table reports no uncertainty or raw episode counts.
Go to primary source ↓e-discussionPDF p. 7, Section 5, Generalization and Modular pipeline vs end-to-end learning; p. 8, RL discussion and Infraction analysis
Authors attribute new-town failures to transfer and perception issues, discuss RL compute and tuning limitations, and note that RL lacks IL's augmentation/dropout. These are explanations, not controlled ablations.
Go to primary source ↓e-infractionsPDF p. 8, Table 2, Opposite lane, Sidewalk, Collision-car and Collision-pedestrian rows; Infraction analysis
Table reports kilometers between infractions. Training car-collision distances are MP/IL/RL 16.4/3.26/0.58; opposite-lane values 10.2/33.4/0.18. Pedestrian values are 18.9/6.35/17.8 in training and 1.40/1.88/2.55 in new town. New-weather IL sidewalk value is printed >57.
Go to primary source ↓e-infraction-definitionPDF p. 16, S.3.1, five infraction definitions and final paragraph
Wrong-lane and sidewalk infractions exceed 30% of vehicle footprint. Collisions are contact events. Continuing violations are counted in intervals of at most two seconds.
Go to primary source ↓e-conclusionPDF p. 8, Section 6
The conclusion describes the simulator and assets as open and states they will be released at the project site; it provides no physical deployment evaluation.
Go to primary source ↓8.5 Primary sources
CARLA: An Open Urban Driving Simulator ↗
PDF · 8,114 extracted words
Source fingerprint
366f3b9a64d8b34c4f9eb698a4dc865958a9219ddf872e8c497843a9ae01d650