Physical AI & The Creation of The Physics Endpoint
I believe that everyone should be able to effectively communicate with robots, and in order to level the playing field to this effect I believe the solutons need to be universal.
Making Robots Easy to Understand
A Framework for Distributed Physical Intelligence
Fletcher Hillier
January 2026
Abstract
This paper presents a framework for democratizing robotics through input normalization—a method of standardizing spatial perception that separates geometric understanding from semantic processing. By combining low-cost stereoscopic sensors with server-side physics calculations, this approach intends to enable universal access to robotic capability while creating substantial new economic opportunities for AI infrastructure providers. The framework addresses the growing risk of robotic oligopoly by distributing the core enabling technology at an accessible cost, ensuring that the benefits of automation extend to the general population rather than asymmetrically concentrating production capability. This paper provides relatively low resolution technical detail for implementation, deliberately leaving room for diverse solutions and contributions in a growing solution space.
Executive Summary
The robotics industry stands at an inflection point. Market projections indicate growth from $90.2 billion in 2024 to $205.5 billion by 2030—a 15% compound annual growth rate that will reshape global production. The critical question is not whether this transformation will occur, but who will control it and who will benefit from it.
This paper proposes input normalization as the technical foundation for distributed robotic capability. The core insight is that spatial understanding—knowing where objects are in three-dimensional space—can be separated from the complex physics calculations required to interact with those objects. When this separation occurs, the expensive computation moves to scalable server infrastructure while the local sensing hardware becomes radically simplified and inexpensive.
The economic implications are substantial. A functional robotics sensing unit can be constructed for under twenty dollars at retail pricing, compared to thousands of dollars for current LIDAR-based systems. This cost reduction opens robotics participation to individuals, educators, small manufacturers, and developing economies currently priced out of the market.
For AI infrastructure providers, this framework represents a new category of API revenue. Physics calculation endpoints would serve continuous demand from every deployed robot—a market potentially larger than current language and image generation services combined. The infrastructure already exists; what has been missing is a standardized input method to connect it to physical reality.
The strategic rationale for open publication is straightforward: preventing the consolidation of robotic capability in a small number of corporations before that consolidation becomes irreversible. The window for action is measured in years, not decades.
Part I: The Market Opportunity
1.1 The Scale of the Transformation
The global robotics market represents one of the most significant growth opportunities in the current technological landscape. According to GlobalData’s 2025 analysis, the market will more than double from $90.2 billion to $205.5 billion by 2030. However, projections vary across analysts depending on segment definitions and methodology. The following table presents ranges where estimates diverge:
Total Robotics
$90.2B
$110-205B
10-15%
Range reflects industrial vs. service inclusion
Humanoid Robots
$2B
$15-18B
~35-40%
Emerging segment, high variance
Collaborative Robots
$2B
$10-12B
~30-35%
Depends on safety certification pace
Consumer Robotics
$10.92B
$29-40B
18-25%
Grand View: $40B; others lower
Intelligent Robotics
—
—
29.2%
MarketsandMarkets estimate
These figures, while varying in specifics, consistently indicate a fundamental shift in how physical work is performed. Precedence Research projects the consumer robotics segment alone could reach $102.31 billion by 2034, suggesting sustained long-term growth beyond current forecasting horizons. The variance across sources reflects genuine uncertainty about adoption rates, regulatory environments, and technological maturation—but the directional trend is consistent across all credible analyses.
1.2 The Untapped Physics Endpoint
Current AI infrastructure has optimized for three primary categories: language processing, code generation, and image synthesis. These endpoints serve knowledge workers, developers, and creative professionals—a significant but ultimately bounded market segment. The majority of API calls originate from a relatively small population of technical users and the applications they build.
A physics calculation endpoint represents a potentially different scale of opportunity. Consider the use case density:
Language endpoints: used when humans compose text—perhaps dozens of interactions per day for active users, with most of the global population never engaging directly.
Physics endpoints: used whenever a robot interacts with its environment—potentially thousands of calculations per hour, per robot, continuously. Every household task, every factory operation, every navigation decision could generate API demand.
The installed base comparison is instructive. ChatGPT reached 100 million users within two months of launch—an unprecedented adoption rate for a software service. A physics endpoint serving robotics could eventually reach billions of deployed units, each generating continuous calculation requests. If this adoption materializes, the aggregate demand could exceed language processing by orders of magnitude.
1.3 The Infrastructure Alignment
The computational infrastructure required for physics calculations already exists and is actively expanding. GPU clusters optimized for parallel processing—originally built for graphics rendering, then repurposed for AI training—are ideally suited for physics simulation.
NVIDIA’s public statements indicate aggressive infrastructure expansion. The company’s robotics simulation platform, Isaac Sim, already demonstrates the viability of GPU-accelerated physics for robotic applications. Their Isaac Lab framework achieves approximately 9,200 samples per second on an RTX 4080 for training tasks across 4,096 parallel environments—a 16× speedup over CPU-based alternatives.
The key insight from Columbia University’s Creative Machines Lab reinforces this approach: “Internal computational models of physical bodies are fundamental to the ability of robots and animals alike to plan and control their actions. These ‘self-models’ allow robots to consider outcomes of multiple possible future actions, without trying them out in physical reality.”
This is precisely what a physics endpoint provides: the ability to simulate actions before executing them, leveraging computational resources that would be impractical to deploy on every individual robot.
1.4 The Cost Reduction Opportunity
Current robotics sensing systems impose substantial cost barriers. Historically, a Velodyne Ultra Puck VLP-32C LIDAR sensor carried a price point of approximately $8,000, though recent industry consolidation (including the Velodyne-Ouster merger) and increased competition have begun to reduce prices in some market segments. Entry-level rotating LIDAR units start at $69-100, with precision units like the Garmin LIDAR-Lite v3HP at approximately $150. These costs, while declining, still reflect the complexity of generating dense three-dimensional mesh data—complexity that input normalization renders unnecessary.
The proposed approach reduces sensing requirements to stereoscopic triangulation using commodity components:
HC-SR04 Ultrasonic Sensor
$1-5
<$1
VL53L0X Time-of-Flight
$5-11
$2-4
ESP32 Microcontroller
$3-12
$2-3
NEMA 17 Stepper Motor
$8-15
$4-8
OV7670 Camera Module
$5-8
$2-4
A complete sensing unit incorporating dual distance sensors, a microcontroller, and basic motor actuation can be assembled for $8-18 at retail pricing. At production volumes, this cost drops to $5-10. Even accounting for continued LIDAR price reductions, the contrast with mesh-based sensing systems represents a cost reduction of one to two orders of magnitude—sufficient to fundamentally change accessibility economics.
Part II: Technical Framework
2.1 The Reference Line Method: Mathematical Foundation
The foundation of input normalization is stereoscopic triangulation—a technique that mirrors biological binocular vision. Two sensors at a known baseline distance (B) each observe the same point in space. The angular displacement between these observations (the disparity, d) combined with the optical characteristics of the sensors (focal length, f) determines the depth (Z) to the observed point.
The governing equation is:
Z = (f × B) / d
Where:
Z = depth to the observed point
f = focal length of the sensors
B = baseline distance between sensors (fixed, known)
d = disparity (horizontal pixel displacement between left and right observations)
This relationship is inverse: objects close to the sensors produce large disparities and therefore precise depth measurements, while distant objects approach zero disparity as they approach optical infinity.
Carnegie Mellon University’s Computer Vision curriculum (16-385) establishes the mathematical rigor: “Each 2D-to-3D correspondence provides two equations. With two views, the system becomes overdetermined and can be solved via singular value decomposition.” This overdetermined nature provides inherent error correction—the mathematics produce the best-fit solution even when individual measurements contain noise.
Practical calculation examples are provided in the original paper illustrating distance-resolution tradeoffs at specific baselines and focal lengths; these highlight how disparity resolution limits depth precision at range while yielding fine resolution at close range suitable for manipulation.
2.2 Multi-Sensor Redundancy: Aerospace Heritage
Real-world sensors produce imperfect data. Environmental interference, component degradation, and transient faults corrupt individual measurements. The solution draws from aerospace engineering: redundant sensors with voting or averaging to reject outliers.
NASA’s Triple Modular Redundancy (TMR) approach has documented heritage spanning six decades. The Saturn V Launch Vehicle Digital Computer (LVDC), designed by IBM between 1967 and 1973, implemented triple-redundant logic with majority voting at each pipeline stage. NASA technical documentation reports reliability exceeding 99.6% over 250 hours of continuous operation.
The mathematical foundation follows the reliability formula:
R_TMR = R_v (3 R_m^2 − 2 R_m^3)
When component reliability exceeds 0.5, TMR produces net reliability improvement. At R_m = 0.9:
R_TMR = 1.0 × (3×0.81 − 2×0.729) = 2.43 − 1.458 = 0.972
System reliability (97.2%) exceeds individual component reliability (90%) despite requiring three components.
For sensor applications, averaging reduces noise by √N. Voting or median selection eliminates gross outliers entirely. This suggests each “eye” in the proposed system should incorporate multiple sensor modalities—infrared, ultrasonic, and time-of-flight—providing independent measurements that can be fused for robust results before data leaves the local unit.
2.3 Data Efficiency: The Bandwidth Advantage
A fundamental distinction between mesh-based sensing (LIDAR) and coordinate-based sensing (reference line method) lies in data volume. This has cascading implications for bandwidth, processing requirements, and system architecture.
LIDAR data example: a Velodyne VLP-32C produces ~1.2M points/s. At 28 bytes/point, bandwidth is ~33.6 MB/s; a ten-second scan produces ~364 MB of raw data.
Reference line method: discrete points of interest rather than dense meshes. A single 3D coordinate requires ~12 bytes; adding timestamp and confidence brings this to ~20 bytes/point. For navigation, 100–200 reference points at 10 Hz: ~40 KB/s—an ~840× reduction versus LIDAR.
Processing implications: LIDAR needs graphical rendering pipelines for point-cloud processing; coordinate-based approaches operate in mathematical space—no mesh reconstruction required—reducing compute and latency.
2.4 The Layered Response Architecture
Network latency presents an obvious challenge for server-dependent robotics. Round-trip times of 20–100 ms are incompatible with real-time motor control requiring sub-millisecond response. The solution is a layered system that matches response requirements to processing location.
Immediate Safety (Onboard, <1ms response)
Collision detection and emergency stops execute locally with zero network dependency. This layer operates continuously, monitoring proximity sensors and enforcing hard boundaries. Even consumer devices like the Roomba implement this basic safety layer—the physics endpoint does not replace it.
Pattern Recognition (Onboard, <10ms response)
Previously learned patterns—object classifications, environmental features, routine movements—are recognized locally using lightweight models. An apple-sorting robot does not re-analyze “apple” with every encounter; it matches against stored patterns developed during training. This layer requires modest local processing capability.
Learning and Optimization (Batch, minutes to hours)
Retroactive analysis examines outcomes against predictions. Successful action pathways are reinforced; unsuccessful ones are diminished. This layer operates without time pressure, processing accumulated experience data to improve future performance.
FogROS2 research demonstrates that properly layered architectures can achieve collision avoidance and reduced failure rates even under network congestion.
2.5 Self-Discovery: The Robot Learns Its Own Body
Input normalization enables robots to discover their own kinematics and capabilities through perception rather than relying on brittle pre-programmed models. The robot observes its own body as objects in the physics environment and builds a self-model.
This process resembles infant motor development. The physics endpoint can guide exploration, prevent dangerous movements, and accelerate learning. Modularity and repair become simpler because a robot that rediscovers its configuration adapts without reprogramming.
2.6 The Motor Calibration Breakthrough
When spatial perception observes motor outcomes directly, actuator precision becomes less critical. The system empirically learns transfer functions: given a control input, observe the resulting position change and update models accordingly. Visual servoing literature (IBVS) documents robustness to calibration errors when closed-loop visual feedback is used.
Economic implications: low-cost, salvaged, or worn motors become viable components; manufacturing tolerances can be relaxed; education and hobbyist robotics benefit.
2.7 The Physics Teacher Model
The endpoint functions as a teacher, instructing onboard learning systems about feasible actions and constraints.
Low-capability units depend on the teacher continuously (higher API usage).
High-capability units retain learned behaviors locally and consult the endpoint for novel situations.
Factories initially pay to learn basic operations and later pay to optimize.
This framing clarifies business models: ongoing optimization and improvement become revenue drivers for endpoint providers. Sim-to-real research supports high success rates from simulation-trained policies.
Part III: The Eyeball Unit
3.1 Physical Design Specifications
The standardized sensing unit—the "eyeball bar"—consolidates perception, processing, and communication into a compact package.
Dimensions
50mm × 50mm × 200mm
Baseline Distance
150mm (fixed, mechanically guaranteed)
Sensor Complement
2× IR distance, 2× ultrasonic, 2× ToF per eye
Actuation
2× NEMA 14/17 steppers per eye (pan/tilt)
Processing
ESP32-S3 or equivalent (Wi‑Fi, BLE integrated)
Expansion
GPIO breakout for motor control (8–16 channels)
Power
5V/2A via USB‑C or 12V barrel jack
Baseline rigidity is critical: the distance between sensor clusters must remain constant during scanning. A rigid central bar with gimbaled sensor mounts is recommended.
3.2 Multi-Pass Scanning Strategy
Environmental perception operates in progressive resolution passes to balance awareness and load:
Pass 1 — Collision Avoidance (continuous, low resolution): 20–50 reference points at 10–20 Hz.
Pass 2 — Navigation Mapping (periodic, medium resolution): 100–200 reference points before movement.
Pass 3 — Manipulation Preparation (on-demand, high resolution): 500+ points concentrated on target object and surroundings.
Pass 4 — Active Monitoring (during task, adaptive resolution): adjust temporal/spatial resolution based on task dynamics.
Entry-level units perform fewer, lower-resolution passes; premium units perform more frequent, higher-resolution passes and more onboard processing.
3.3 Product Tier Structure
A product ladder spanning educational to industrial applications:
Entry Tier ($15–25 retail): ESP32, dual HC‑SR04, stationary mount, Wi‑Fi only, endpoint-dependent.
Standard Tier ($75–125 retail): ESP32‑S3, VL53L1X ToF, camera, NEMA 14 motors, 8‑channel motor control, basic onboard pattern recognition.
Professional Tier ($200–350 retail): Coral TPU, redundant sensor arrays, high‑precision ToF, NEMA 17 motors, 16‑channel control with current sensing, IP54, substantial onboard learning.
All tiers connect to the same physics endpoints; tiers represent autonomy/cost tradeoffs, not incompatible ecosystems.
Part IV: Strategic and Economic Analysis
4.1 The Consolidation Risk
Without intervention, robotics development can concentrate in a few dominant companies—repeating patterns seen in search, social networks, cloud infrastructure, and mobile OS markets. When robots perform large portions of productive work, control of robotic capability becomes control of economic production. Democratizing access to foundational perception and physics calculation reduces the risk that automation benefits concentrate narrowly.
4.2 The Open Distribution Rationale
Publishing this framework openly aims to establish prior art and prevent exclusive capture via patents. Open distribution expands the addressable market for physics endpoints, benefits hardware manufacturers through standardization-driven volume, and enables robot builders to innovate on hardware without reimplementing core perception and physics stacks. End users benefit from competition and reduced lock-in.
4.3 Market Size Projections
Order-of-magnitude illustrative projections:
Factory robotics (conservative): up to 28 trillion endpoint calls annually at scale; potential ~$28B/yr at $0.001 per 1,000 calls assumption.
Consumer robotics (optimistic): 50M units × 182,500 calls/year = 9.1 trillion calls; potential ~$9B/yr at same pricing.
Combined illustrative TAM: ~$37B annually by 2030 under stated assumptions (excludes AVs, drones, agriculture).
Projections depend on standardization, latency solutions, safety frameworks, and adoption rates.
4.4 Alignment with Major Players
NVIDIA: invested in Isaac Sim/Omniverse; physics endpoints leverage GPU investments.
Anthropic/OpenAI/AI infra companies: can extend API portfolios to physics calculations.
Humanoid manufacturers: can reduce perception development costs and focus on hardware differentiation.
Educational institutions: benefit from lower-cost hardware enabling broader participation.
Part V: Implementation Domains
5.1 Factory Automation
Controlled, largely static environments are high-value early applications. A layered architecture applies: server-modeled environment, local pattern recognition, onboard safety, server consultation for novel or optimized actions, batch optimization for throughput improvements. Economic value derives from reducing bottlenecks and downtime.
5.2 Consumer Home Robotics
Home environments are variable but have lower precision needs for many tasks. Subscription economics align naturally—monthly fees for physics processing and continuous improvement. Projected consumer pricing (illustrative):
Entry (basic chores)
$400–600
$15–25
$580–900
Standard (full household)
$1,500–2,500
$30–50
$1,860–3,100
Premium (comprehensive)
$3,000–5,000
$50–100
$3,600–6,200
These points compare favorably to existing high-end consumer robots while offering broader capability.
5.3 Education and Research
Input normalization lets students interact directly with hardware while perception and physics are abstracted. Competitive programs (e.g., FIRST) could standardize on eyeball sensors, shifting focus to mechanical creativity. Low-cost hardware amplifies accessibility for institutions in developing regions.
5.4 Assistive Technology and Aging Populations
Affordability, adaptability, safety through endpoint moderation, and subscription sustainability make the framework well-suited to assistive robots for elderly and disabled populations.
5.5 Remote and Hazardous Environments
Applications include nuclear maintenance, deep-sea exploration, disaster response, and planetary operations. Latency constraints necessitate high‑capability onboard autonomy; the physics endpoint provides training, optimization, and post‑mission analysis.
5.6 Workforce Transition Considerations
Automation displaces routine jobs. Democratizing robotics shifts the distribution of productivity gains toward broader populations rather than consolidating them within a few firms. Policy and retraining programs remain necessary to manage workforce transitions.
Part VI: Safety, Constraints, and Governance
6.1 Operational Boundaries
Robots must not extrapolate beyond authorized training. The physics endpoint maintains authoritative knowledge of each robot’s training scope; out-of-scope requests trigger escalation rather than improvised action.
Exploration during self‑discovery must respect hard limits supplied by the endpoint (e.g., voltage, joint angle caps).
6.2 Human-in-the-Loop Requirements
Situations requiring human judgment include unresponsive motors, unmatched environments, out-of-scope task requests, and anomalous sensor readings. Escalation paths must be clear; human operators must be able to assume immediate control. Systems must fail safe.
6.3 Auditability and Accountability
The distributed architecture creates comprehensive audit trails: input data, calculation requests, responses, execution logs, and outcomes. This supports quality assurance, incident investigation, continuous improvement, and regulatory compliance. Liability allocation across manufacturers, endpoint providers, operators, and owners will require legal frameworks beyond this paper, but the framework provides an evidentiary foundation.
6.4 Endpoint Moderation Responsibility
Centralized physics calculations create a chokepoint where policy can be enforced. Endpoint providers must define permissible calculations and refuse to simulate harmful actions (e.g., trajectories intended to harm people, weapon optimization, disabling safety systems). Refusal should occur before computation.
Important moderation practices:
Define prohibited calculation categories
Refuse to simulate harmful actions
Maintain audit logs of rejected and high-risk requests
Integrate human review for edge cases
Provide termination authority and documented appeal processes
This dependency can be a safety advantage: robots reliant on moderated endpoints operate within enforced constraints.
6.5 The Multi-Provider Ecosystem
A multi-provider ecosystem prevents single-vendor lock-in and related risks (policy changes, outages, hostile acquisition). Components:
General-purpose providers (broad capabilities)
Specialized providers (vertical optimization)
Regional providers (low latency, regulatory compliance)
Open-source/community endpoints (education/research)
Interoperability standards are required for provider switching, common input/output formats, capability advertisement, and authentication protocols. Critical deployments should maintain primary/secondary/tertiary provider relationships for failover. Standards bodies (IEEE, ISO) can provide formalization.
6.6 Physics-Verified Accountability
Physical reality provides ground truth: actions produce observable outcomes that can be automatically compared to predictions. This enables:
Automatic outcome verification (did the gripper reach its target?)
Detection of systematic prediction errors or anomaly patterns
Continuous fleet-scale model improvement from real outcomes
Regulatory approaches focused on continuous monitoring and mandatory reporting rather than purely pre-deployment certification
Verification supports both safety monitoring and model refinement.
6.7 Preventing Misuse
Layered defenses:
Endpoint policy: refuse prohibited calculations and log attempts.
Multi-provider oversight: competitive pressure and redundancy.
Physics verification: anomalies produce detectable evidence.
Human review: specialists handle edge cases.
These layers raise the barrier to misuse and make detection and attribution more feasible. Fully self-contained autonomous systems remain a risk vector for determined malicious actors.
6.8 Toward Open Standards
Key standardization areas:
Input formats (coordinates, reference frames, sensor metadata)
Output formats (responses, confidence/uncertainty, guidance)
Protocols (authentication, session management, failover)
Capability advertisement and benchmarking
Safety certification attestation mechanisms
Standards should emerge from open collaborative processes to avoid entrenching dominant players. This paper invites participation in those efforts.
Conclusion
Input normalization is a technically sound, economically viable, and strategically important approach to democratizing robotics. It integrates stereoscopic triangulation, redundant sensing, and distributed computation to produce an architecture that lowers hardware costs, enables broad participation, and creates a new class of physics endpoint APIs with substantial market potential.
The strategic imperative is to prevent concentrated control of robotic capability. Open distribution, multi-provider ecosystems, standardization, and endpoint moderation collectively reduce consolidation risk while enabling innovation and competition.
The framework is intentionally implementation‑friendly but non‑prescriptive. The invitation is explicit: build on this framework, improve it, compete with it, deploy it. The goal is broad participation—this opportunity belongs to everyone willing to pursue it.
About the Author
Fletcher Hillier is an entrepreneur and technologist based in Canada, working at the intersection of AI systems, automation, and economic accessibility. This paper represents independent research conducted without institutional affiliation or external funding.
Correspondence: Published for open distribution.
License: This paper is released for unrestricted distribution. Readers are encouraged to share, adapt, and build upon this work. Attribution is appreciated but not required. The explicit intent is to establish prior art preventing exclusive patent claims on the described framework.
References
Ambarella. (2024). A Closer Look at LiDAR and Stereovision: Performance Analysis and Comparison. Technical Analysis.
Carnegie Mellon University. (2017). Computer Vision: Triangulation (16‑385 Course Materials). Pittsburgh, PA.
Chaumette, F., & Hutchinson, S. (2006). Visual Servo Control Part I: Basic Approaches. IEEE Robotics and Automation Magazine, 13(4), 82–90.
Chaumette, F., & Hutchinson, S. (2007). Visual Servo Control Part II: Advanced Approaches. IEEE Robotics and Automation Magazine, 14(1), 109–118.
Chen, B., Kwiatkowski, R., Vondrick, C., & Lipson, H. (2022). Full‑body visual self‑modeling of robot morphologies. Science Robotics, 7(68).
Columbia University Creative Machines Lab. (2022). A Robot Learns to Imagine Itself. Press Release and Technical Documentation.
FogROS2 Development Team. (2023). FogROS2‑LS: A Location‑Independent Framework for Cloud Robotics. UC Berkeley.
GlobalData. (2025). Global Robotics Market Forecast 2024–2030: Analysis and Projections.
Grand View Research. (2024). Consumer Robotics Market Size, Share & Trends Analysis Report, 2024–2030.
Hutchinson, S., Hager, G. D., & Corke, P. I. (1996). A Tutorial on Visual Servo Control. IEEE Transactions on Robotics and Automation, 12(5), 651–670.
IBM Corporation. (1967–1973). Saturn V Launch Vehicle Digital Computer Technical Documentation. Prepared for NASA.
International Organization for Standardization. (2011). ISO 10218‑1:2011 Robots and robotic devices — Safety requirements for industrial robots.
International Organization for Standardization. (2014). ISO 13482:2014 Robots and robotic devices — Safety requirements for personal care robots.
International Organization for Standardization. (2016). ISO/TS 15066:2016 Robots and robotic devices — Collaborative robots.
MarketsandMarkets. (2024). Intelligent Robotics Market Size, Share, Trends and Growth Drivers 2024–2032.
Mittal, M., et al. (2025). Isaac Lab: A Unified Framework for Robot Learning Built on NVIDIA Isaac Sim. arXiv preprint.
National Academies Press. (1996). Statistical Software Engineering: Case Study: NASA Space Shuttle Flight Control Software. Washington, DC.
Niryo. (2024). Accessibility to Robotics, Five Years Later: Progress and Remaining Challenges.
NVIDIA Corporation. (2024). Training Sim‑to‑Real Transferable Robotic Assembly Skills over Diverse Geometries. NVIDIA Technical Blog.
Precedence Research. (2024). Consumer Robotics Market Size and Growth Projections to 2034.
Velodyne LiDAR. (2018). VLP‑16 Puck and VLP‑32C Ultra Puck Datasheets.
Zhao, W., Queralta, J. P., & Westerlund, T. (2020). Sim‑to‑Real Transfer in Deep Reinforcement Learning for Robotics: A Survey. arXiv:2009.13303.
Last updated