Pith. sign in

REVIEW 3 major objections 6 minor 55 references

Ark: An Open-source Python-based Framework for Robot Learning

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Ark's central claim is that one Gym-style framework can carry a robot-learning pipeline from data collection to real deployment by flipping a single configuration flag.

desk verdict Ark is a solid Python-first robot-learning framework whose headline sim-to-real switch is demonstrated only with a hand-crafted command-echo policy, so the load-bearing claim about learned policies remains untested. read the letter →

arxiv 2506.21628 v2 pith:TVDENCDF submitted 2025-06-24 cs.RO cs.AIcs.LG

Magnus Dierking , Christopher E. Mower , Sarthak Das , Huang Helong , Jiacheng Qiu , Cody Reading , Wei Chen , Huidong Liang
show 5 more authors
Huang Guowei Jan Peters Quan Xingyue Jun Wang Haitham Bou-Ammar
This is my paper · ORCID
classification cs.ROcs.AIcs.LG
keywords robotlearningframeworkimitationsim-to-realtransferGym-styleinterfacepublisher-subscribermiddlewaremanipulationmobilenavigationembodiedAI
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Ark is a software framework whose central claim is that robot learning can be made as approachable as mainstream machine learning by giving every robot task the same Gym-style interface and routing all communication through a distributed message network. It aims to let a user collect demonstrations, train imitation-learning policies such as ACT and Diffusion Policy, and deploy those policies on a simulated or physical robot without changing code, by flipping one configuration flag. If this works, robotics researchers and engineers no longer need separate stacks for simulation, data collection, and real deployment, which is currently a major bottleneck in embodied AI. The paper supports the claim with several worked use cases spanning manipulation, mobile navigation, and LLM-driven control.

What carries the argument

The load-bearing mechanism is the observation-and-action channel abstraction running on Ark's distributed publisher-subscriber network, built on the LCM messaging library. Each robot, sensor, simulator, or policy is a node; observation and action spaces are dictionaries mapping channel names to message types; and the sim-real switch reroutes those channels to simulator backend nodes or to hardware drivers while preserving names and schemas. The channel abstraction is what lets one policy pipeline operate in both domains without modification, and the YAML configuration file is what lets users change the routing by toggling sim.

What would settle it

Run a policy trained entirely in a supported simulator on a physical robot through Ark with only the sim flag changed and measure task success over many trials; if success rates drop sharply relative to simulation because of latency, noise, or calibration differences in the message streams, the one-flag seamless-switch claim is falsified.

Watch

Extended reading notes

Core claim

On its own terms, Ark's contribution is an architecture-level claim: when every sensor and actuator, whether real or simulated, is exposed as an independent node publishing and subscribing to typed message channels, and the environment's observation and action spaces are defined as collections of those channels, then simulation and reality become interchangeable backends. The paper demonstrates this with a pick-and-place task in which a hand-crafted expert policy runs unchanged on a fixed-base robotic arm with only the configuration variable sim changed from True to False. It further shows the same channel definitions supporting kinesthetic and teleoperated data collection, diffusion policy and ACT training, SLAM and A* navigation, and an LLM-based policy selector, all within one framework.

Load-bearing premise

The load-bearing premise is that the message channels and data formats that simulated drivers produce are behaviorally equivalent to those of real hardware drivers, so that toggling sim=True/False yields the same policy behavior with zero code changes.

Editorial extensions

If this is right

  • A user who writes one policy pipeline in Ark can run it in a supported simulator and on supported physical hardware without rewriting code, because the same channels carry both simulated and real data.
  • The same observation and action configuration used for deployment is used for extracting logged demonstrations into CSV, so imitation-learning data collection and training align with deployment by construction.
  • Hardware vendors and labs can expose new robots by implementing driver subclasses, making new embodiments available to the same policy code and data tools.
  • Existing setups built on the widely used robotics middleware can be connected through Ark's bridge driver, so teams can migrate incrementally rather than rebuilding their stacks.
  • The reusable nodes for control, SLAM, planning, and visualization allow mobile navigation and embodied-AI workflows to be assembled from the same building blocks as manipulation pipelines.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The strongest version of the one-flag claim is not yet stress-tested in the paper: the sim-real demonstration uses a hand-crafted expert policy on a single pick-and-place task, so the same guarantee for learned policies under real-world latency, sensor noise, and calibration error remains an inference from the architecture rather than an established result.
  • If the channel-equivalence assumption holds across embodiments, Ark could serve as a common benchmark harness for sim-to-real transfer, because the same message schema would make policy success rates directly comparable across simulators and robots.
  • An immediate testable extension would be to deploy a policy trained entirely in simulation on each supported robot with only the sim flag changed and report closed-loop success rates, which would quantify any residual sim-real gap that the channel abstraction alone cannot remove.
  • The paper states that the current messaging backend is not well suited for distributed training, so the long-term claim that Ark accelerates embodied-AI research may depend on replacing or augmenting the communication layer.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper presents Ark, an open-source, Python-first robot learning framework with a Gym-style environment interface, a publisher-subscriber network built on LCM, interchangeable drivers for simulated and physical hardware, a ROS bridge, and tools for data collection, visualization, and debugging. After describing the architecture and design principles, it reports case studies that include a sim-real switch on a ViperX pick-and-place task, imitation learning with Diffusion Policy and ACT on real robots, SLAM and navigation on a Husky, and an LLM-based board-game agent using DeepSeek-R1. The central claim is that a single configuration flag (sim = True/False) allows the same observation/action channel definitions and policy pipeline to run in both simulation and on physical hardware without code changes. The paper is qualitative and descriptive; it does not report quantitative measurements of latency, control frequency, or task success rates.

Significance. If the central sim-real claim were convincingly established, Ark would be a useful community contribution: it combines a familiar Gym-style interface, Python-first workflows, modular driver and networking abstractions, and integrated data-collection and debugging tools, which could lower the engineering barrier for robot learning research. The open-source code, documented case studies, and ROS interoperability are concrete assets. However, the load-bearing evidence for the seamless sim-real switch is currently missing, so the significance of the framework beyond existing tools such as LeRobot and PyRobot is not yet demonstrated.

major comments (3)
  1. [Use Cases, 'Switching Between Simulation and Reality' (Fig. 4)] The only direct sim-real comparison uses an observation space defined as 'the current joint position command' and an action space of 'goal robot joint velocity command' with a hand-crafted expert policy. Because the observation is the command just sent, it is identical by construction in simulation and reality; this does not exercise encoder feedback, camera images, actuation latency, or controller dynamics. The central claim that a single flag enables a policy pipeline to operate unchanged in both domains therefore requires a demonstration in which a learned policy (e.g., Diffusion Policy or ACT) is trained in a supported simulator and then deployed on the physical robot by only toggling sim, with task success rates or quantitative trajectory metrics reported.
  2. [Use Cases, 'Imitation Learning'] The imitation-learning demonstrations (Diffusion Policy on the ViperX pushing task and ACT on the OpenPyRo-A1 cloth manipulation and handover tasks) are presented as real-hardware deployments only; no experiment trains a policy in PyBullet or MuJoCo and then deploys it on the physical system via the sim flag. The Future Work section also acknowledges that the simulator stack currently lacks domain randomization and differentiable physics, which further weakens the implicit claim that simulation-trained policies can transfer seamlessly. Please either provide such a cross-domain experiment or explicitly scope the claim to code and interface reuse rather than learned-policy sim-to-real transfer.
  3. [Table 1 and overall evaluation] The paper contains no quantitative evaluation of the framework's core performance promises: there are no measurements of control frequency, end-to-end communication latency, task success rates, or user effort, and Table 1 is a self-reported feature matrix with no stated evaluation protocol or reproducibility basis. Since the abstract and introduction emphasize 'lightweight' communication, 'real-time performance when needed', and 'seamless switching', these claims should be supported by at least basic benchmarks (e.g., publisher-subscriber latency, control-loop rates in simulation and reality, and success rates for the pick-and-place, pushing, cloth manipulation, and handover tasks).
minor comments (6)
  1. [Throughout] There are numerous typographical and formatting issues (e.g., 'T o' for 'To', 'recieved' for 'received', 'it' s' for 'it's'); the manuscript would benefit from a careful proofreading pass.
  2. [Throughout] The framework name is inconsistently rendered as 'ARK' and 'Ark'; please pick one capitalization and use it consistently.
  3. [Use Cases] The statement 'All code to reproduce these use-cases will be made available' is a promise rather than a current artifact; please include a link to a tagged release or tested version of the repository with the submission.
  4. [Embodied AI] The LLM tournament reports only overall win rates (Qwen 2.5: 26.6%, Llama 3: 30.0%, DeepSeek-R1: 43.3%) with no number of games, variance, or statistical test, so the 'highest win-rate' claim is not yet substantiated.
  5. [Framework Overview, 'Simulator backend'] Minor wording issue: 'IssacSim' should be 'IsaacSim'.
  6. [Framework Overview, 'Ark Network'] The description of message channels, services, and the registry would be easier to follow with a small code snippet showing how a user defines an observation/action space and launches a node.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: Ark is a software-architecture description; its sim-real switch is an implementation property, not a fitted prediction, and self-citations are not load-bearing.

full rationale

This is a software-framework paper, not a mathematical derivation, so the usual fitted-parameter or self-citation circularity patterns do not apply. No parameter is fitted to data and then reported as a prediction. The central sim-real switching claim is realized through architecture: the paper states that Ark 'ensures consistent message schemas, channel names, and execution flow' when switching between simulated and physical systems, which is an implementation guarantee rather than an empirically derived result. The paper does not invoke a uniqueness theorem, an ansatz, or a prior author result to rule out alternatives. The self-citations, e.g., OpenPyRo-A1 [48] and ROS-LLM [31], are used as hardware-platform and related-work references and do not carry the load-bearing argument about Ark's capabilities. The main weakness is evidential rather than circular: the Figure 4 sim-real demonstration uses a hand-crafted policy whose observation is 'the current joint position command,' so it does not validate the zero-code switch for learned vision-based policies under real-world sensor noise and latency. That is a support gap, not a circular derivation.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

The paper is a software framework description, so there are no fitted free parameters or invented physical entities. The two domain assumptions are the main premises about Python and LCM that the framework's design relies on. The central sim-real equivalence is the claim being evaluated, not an input axiom.

assumptions (2)
  • domain assumption Python-centric interfaces are easier for machine learning researchers to adopt than C/C++ and are sufficient for most robotics workflows with optional C++ bindings.
    This underlies design principle D3 and the framing of C/C++ as a barrier. The paper does not provide empirical evidence (e.g., user studies) comparing development speed or ease of use.
  • domain assumption LCM (Lightweight Communications and Marshalling) provides low-latency, reliable communication sufficient for controlling robots and streaming sensor data, while remaining simple to use.
    The paper selects LCM as the networking backend but provides no latency, throughput, or jitter benchmarks comparing it to ROS or other middleware. The authors acknowledge that LCM may not support future distributed training needs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ark: An Open-source Python-based Framework for Robot Learning." pith.science (2026). https://pith.science/paper/TVDENCDF

@misc{pith2026250621628,
  author       = {Pith},
  title        = {Pith review of: Ark: An Open-source Python-based Framework for Robot Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TVDENCDF}},
  note         = {Machine review of arXiv:2506.21628}
}
read the original abstract

Robotics has made remarkable hardware strides-from DARPA's Urban and Robotics Challenges to the first humanoid-robot kickboxing tournament-yet commercial autonomy still lags behind progress in machine learning. A major bottleneck is software: current robot stacks demand steep learning curves, low-level C/C++ expertise, fragmented tooling, and intricate hardware integration, in stark contrast to the Python-centric, well-documented ecosystems that propelled modern AI. We introduce ARK, an open-source, Python-first robotics framework designed to close that gap. ARK presents a Gym-style environment interface that allows users to collect data, preprocess it, and train policies using state-of-the-art imitation-learning algorithms (e.g., ACT, Diffusion Policy) while seamlessly toggling between high-fidelity simulation and physical robots. A lightweight client-server architecture provides networked publisher-subscriber communication, and optional C/C++ bindings ensure real-time performance when needed. ARK ships with reusable modules for control, SLAM, motion planning, system identification, and visualization, along with native ROS interoperability. Comprehensive documentation and case studies-from manipulation to mobile navigation-demonstrate rapid prototyping, effortless hardware swapping, and end-to-end pipelines that rival the convenience of mainstream machine-learning workflows. By unifying robotics and AI practices under a common Python umbrella, ARK lowers entry barriers and accelerates research and commercial deployment of autonomous robots.

Figures

Figures reproduced from arXiv: 2506.21628 by the authors.

Figure 1
Figure 1. Ark uses a unified configuration file to define action and observation channels, which are then instantiated within a distributed node-based network. This architecture supports both real and simulated hardware through interchangeable drivers and identical communication interfaces. The Ark Registry manages active nodes, while each component (e.g., sensors, actuators, policies) operates as an independent process. As a… view at source ↗
Figure 2
Figure 2. Technical diagram illustrating how Ark uses a unified configuration file to instanti￾ate a distributed simulated system that mirrors real-world deployments. The YAML-based configuration specifies robots, sensors, environments, and networking parameters, which the Ark Simulator parses to launch corresponding simulated nodes. Each component, such as robot controllers, cameras, and sensor emulators, runs as an independ… view at source ↗
Figure 3
Figure 3. Graphical debugging tools provided by Ark. Ark Graph displays active nodes and their communication channels for network analysis. Ark Viewer renders live image streams to support camera calibration and inspection. Ark Plot visualizes real-time numerical data on any channel, aiding in system monitoring and debugging. Ark Plot Ark Plot is a real-time plotting tool for visualizing numeric data on Ark message channels. … view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Seamless sim-real transition in Ark enabled by consistent observation and action space definitions. The environment configuration specifies sensor inputs (e.g., joint states, images) and actuator outputs (e.g., joint commands), which remain identical across both simula…
Figure 5
Figure 5. Figure 5: lcm-logger enables efficient data collection for imitation learning by recording demonstrations from a variety of control interfaces, including kinesthetic teaching, VR teleoperation, and gamepad input (left). Each demonstrations is saved as a separate CSV file (right)…
Figure 6
Figure 6. Figure 6: Sequential snapshots of a Viper X 300 s arm executing a learned diffusion policy. The line overlay shows the sampled action trajectories converging toward the target object, while the arm autonomously refines its actions at each timestep. Teleoperation Ark also support…
Figure 7
Figure 7. Figure 7: Execution of an ACT-based policy on the OpenPyro humanoid robot for two dis￾tinct tasks: cloth manipulation (top row) and object handover (bottom row). The policy produces precise, contact-rich behaviors that enable the robot to flatten a shirt and place a banana into …
Figure 8
Figure 8. Figure 8: Using Ark’s integrated SLAM and visualization tools, a Husky robot equipped with a LiDAR sensor navigates a kitchen environment. The robot first constructs an oc￾cupancy map using FastSLAM (center), facilitated by Ark’s modular data streaming and map-building nodes. Th…
Figure 9
Figure 9. Figure 9: Deep Seek integration with Ark Framework to allow the Viper to play board games right wheel commands using differential drive kinematics. Embodied AI Large language models (LLMs) and vision-language models (VLMs) have been shown to endow robots with impressive reasonin…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 35 canonical work pages

  1. [1]

    Nils J. Nilsson. Shakey the robot. SRI AI Center Technical Note, 1984

    work page 1984

  2. [2]

    Robot manipulators: Mathematics, programming, and control

    Richard Paul. Robot manipulators: Mathematics, programming, and control. In Arti- ficial Intelligence Center, 1972

    work page 1972

  3. [3]

    R. J. Popplestone, A. P . Ambler, and I. Bellos. Rapt: A language for describing assem- blies. Industrial Robot: An International Journal, 5(3):131–137, 1978. ISSN 0143-991X. doi: 10.1108/eb004501. URL https://doi.org/10.1108/eb004501

    work page doi:10.1108/eb004501 1978

  4. [4]

    A robotics toolbox for matlab

    Peter I Corke. A robotics toolbox for matlab. IEEE Robotics & Automation Magazine, 3 (1):24–32, 2002

    work page 2002

  5. [5]

    Vaughan, and Andrew Howard

    Brian Gerkey, Richard T . Vaughan, and Andrew Howard. The player/stage project: T ools for multi-robot and distributed sensor systems. In Proceedings of the Interna- tional Conference on Advanced Robotics (ICAR), 2003

    work page 2003

  6. [6]

    ROS: an open-source robot operating system

    Morgan Quigley, Ken Conley, Brian Gerkey, Josh Faust, Tully Foote, Jeremy Leibs, Rob Wheeler, Andrew Y Ng, et al. ROS: an open-source robot operating system. In ICRA workshop on open source software, volume 3, page 5. Kobe, Japan, 2009

    work page 2009

  7. [7]

    Deep learning

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553): 436–444, 2015. ISSN 1476-4687. doi: 10.1038/nature14539. URL https://doi.org/ 10.1038/nature14539

    doi:10.1038/nature14539 2015

  8. [8]

    Rumelhart, Geoffrey E

    David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. Learning representa- tions by back-propagating errors. Nature, 323(6088):533–536, 1986. ISSN 1476-4687. doi: 10.1038/323533a0

    doi:10.1038/323533a0 1986

Show all 55 references
  1. [9]

    Imagenet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In F . Pereira, C.J. Burges, L. Bottou, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems , volume 25. Curran Associates, Inc., 2012

    2012

  2. [10]

    Long short-term memory

    Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory . Neural Com- putation, 9(8):1735–1780, 11 1997. ISSN 0899-7667. doi: 10.1162/neco.1997.9.8.1735. URL https://doi.org/10.1162/neco.1997.9.8.1735

    1997 doi

  3. [11]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural ...

    2017

  4. [12]

    OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shya- mal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Bal- tescu, Haiming Bao, Mohammad Bavarian, J...

    2024 arXiv

  5. [13]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F . Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei ...

    2025 arXiv

  6. [14]

    O. Khatib. A unified approach for motion and force control of robot manipulators: The operational space formulation. IEEE Journal on Robotics and Automation, 3(1):43–53,

  7. [15]

    A direct method for trajectory opti- mization of rigid bodies through contact

    Michael Posa, Cecilia Cantu, and Russ T edrake. A direct method for trajectory opti- mization of rigid bodies through contact. The International Journal of Robotics Re- search, 33(1):69–81, 2014. doi: 10.1177/0278364913506757. URL https://doi.org/ 10.1177/0278364913506757

    2014 doi

  8. [16]

    Diffusion policy: Visuomotor policy learn- 25 ing via action diffusion

    Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burch- fiel, Russ T edrake, and Shuran Song. Diffusion policy: Visuomotor policy learn- 25 ing via action diffusion. The International Journal of Robotics Research , 0(0): 02783649241273668, 2024. doi: 10.1...

    2024 doi

  9. [17]

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, Jam...

  10. [18]

    A review of robot learn- ing for manipulation: Challenges, representations, and algorithms

    Oliver Kroemer, Scott Niekum, and George Konidaris. A review of robot learn- ing for manipulation: Challenges, representations, and algorithms. Journal of Ma- chine Learning Research, 22(30):1–82, 2021. URL http://jmlr.org/papers/v22/ 19-804.html

    2021

  11. [19]

    Robot learning in the era of foundation models: A survey .Neurocomput- ing, page 129963, 2025

    Xuan Xiao, Jiahang Liu, Zhipeng Wang, Yanmin Zhou, Yong Qi, Shuo Jiang, Bin He, and Qian Cheng. Robot learning in the era of foundation models: A survey .Neurocomput- ing, page 129963, 2025

    2025

  12. [20]

    Robot Program- ming by Demonstration , pages 1371–1394

    Aude Billard, Sylvain Calinon, Rüdiger Dillmann, and Stefan Schaal. Robot Program- ming by Demonstration , pages 1371–1394. Springer Berlin Heidelberg, Berlin, Hei- delberg, 2008. ISBN 978-3-540-30301-5. doi: 10.1007/978-3-540-30301-5_60. URL https://doi.org/10.1007/978-3-540-...

    2008 doi

  13. [21]

    Py- torch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Des- maison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan T ejani, Sasank Chilamkurthy, Benoit Steiner, ...

    2019

  14. [22]

    Pedregosa, G

    F . Pedregosa, G. Varoquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blon- del, P . Prettenhofer, R. Weiss, V . Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay . Scikit-learn: Machine learning in Python. Journal of Machine L...

    2011

  15. [23]

    Openai gym

    Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie T ang, and Wojciech Zaremba. Openai gym. 06 2016. doi: 10.48550/arXiv.1606. 01540

    2016 doi

  16. [24]

    Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Joze- fowicz, Lukasz Kaiser, Manjun...

    2015

  17. [25]

    Dally, Stephen W

    William J. Dally, Stephen W . Keckler, and David B. Kirk. Evolution of the graphics pro- cessing unit (gpu). IEEE Micro, 41(6):42–51, 2021. doi: 10.1109/MM.2021.3113475

    2021

  18. [26]

    cudnn: Efficient primitives for deep learning,

    Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John Tran, Bryan Catanzaro, and Evan Shelhamer. cudnn: Efficient primitives for deep learning,

  19. [27]

    Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Joze- fowicz, Lukasz Kaiser, Manjun...

    2016 arXiv

  20. [28]

    Timperley

    Paulo Canelas, Miguel T avares, Ricardo Cordeiro, Alcides Fonseca, and Christopher S. Timperley . An experience report on challenges in learning the robot operating sys- tem. In Proceedings of the 4th International Workshop on Robotics Software Engi- neering, RoSE ’22, page 33...

    2023

  21. [29]

    It takes a village to build a robot: An empirical study of the ros ecosystem

    Sophia Kolak, Afsoon Afzal, Claire Le Goues, Michael Hilton, and Christopher Steven Timperley . It takes a village to build a robot: An empirical study of the ros ecosystem. In 2020 IEEE International Conference on Software Maintenance and Evolution (ICSME), pages 430–440, 202...

    2020

  22. [30]

    From machine learning to robotics: Challenges and opportunities for embodied intel- ligence, 2021

    Nicholas Roy, Ingmar Posner, Tim Barfoot, Philippe Beaudoin, Yoshua Bengio, Jean- nette Bohg, Oliver Brock, Isabelle Depatie, Dieter Fox, Dan Koditschek, T omas Lozano- Perez, Vikash Mansinghka, Christopher Pal, Blake Richards, Dorsa Sadigh, Stefan Schaal, Gaurav Sukhatme, Den...

    2021 arXiv

  23. [31]

    ROS-LLM: A ROS framework for embodied ai with task feedback and structured reasoning

    Christopher E Mower, Yuhui Wan, Hongzhan Yu, Antoine Grosnit, Jonas Gonzalez- Billandon, Matthieu Zimmer, Jinlong Wang, Xinyu Zhang, Yao Zhao, Anbang Zhai, et al. ROS-LLM: A ROS framework for embodied ai with task feedback and structured reasoning. arXiv preprint arXiv:2406.19...

    2024 arXiv

  24. [32]

    Mower, Haitham Bou-Ammar, and Marco Hutter

    Guowei Lan, Kaixian Qu, René Zurbr ugg, Changan Chen, Christopher E. Mower, Haitham Bou-Ammar, and Marco Hutter. Experience is the best teacher: Grounding vlms for robotics through self-generated memory .Under review, 2025. 27

    2025

  25. [33]

    Hamster: Hierarchical action models for open-world robot manipulation

    Yi Li, Yuquan Deng, Jesse Zhang, Joel Jang, Marius Memmel, Raymond Yu, Caelan Reed Garrett, Fabio Ramos, Dieter Fox, Anqi Li, et al. Hamster: Hierarchical action models for open-world robot manipulation. arXiv preprint arXiv:2502.05485, 2025

    2025 arXiv

  26. [34]

    D. L. Parnas. On the criteria to be used in decomposing systems into modules. Com- mun. ACM, 15(12):1053–1058, December 1972. ISSN 0001-0782. doi: 10.1145/361598. 361623. URL https://doi.org/10.1145/361598.361623

    1972

  27. [35]

    Montemerlo, N

    M. Montemerlo, N. Roy, and S. Thrun. Perspectives on standardization in mobile robot programming: the carnegie mellon navigation (carmen) toolkit. In Proceedings 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2003) (Cat. No.03CH37453), volume 3,...

    2003 doi

  28. [36]

    Robot operating system 2: Design, architecture, and uses in the wild.Science Robotics, 7(66):eabm6074, 2022

    Steven Macenski, Tully Foote, Brian Gerkey, Chris Lalancette, and William Woodall. Robot operating system 2: Design, architecture, and uses in the wild.Science Robotics, 7(66):eabm6074, 2022. doi: 10.1126/scirobotics.abm6074. URL https://www. science.org/doi/abs/10.1126/scirob...

    2022 doi

  29. [37]

    Huang, Edwin Olson, and David C

    Albert S. Huang, Edwin Olson, and David C. Moore. Lcm: Lightweight communica- tions and marshalling. In 2010 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4057–4062, 2010. doi: 10.1109/IROS.2010.5649358

    2010

  30. [38]

    A review of physics simulators for robotic applications

    Jack Collins, Shelvin Chand, Anthony Vanderkop, and David Howard. A review of physics simulators for robotic applications. IEEE Access, 9:51416–51431, 2021. doi: 10.1109/ACCESS.2021.3068769

    2021

  31. [39]

    Lerobot: State-of-the-art machine learning for real-world robotics in pytorch

    Remi Cadene, Simon Alibert, Alexander Soare, Quentin Gallouedec, Adil Zouitine, and Thomas Wolf. Lerobot: State-of-the-art machine learning for real-world robotics in pytorch. https://github.com/huggingface/lerobot, 2024

    2024

  32. [40]

    Pyrobot: An open-source robotics frame- work for research and benchmarking

    Adithyavairavan Murali, T ao Chen, Kalyan Vasudev Alwala, Dhiraj Gandhi, Lerrel Pinto, Saurabh Gupta, and Abhinav Gupta. Pyrobot: An open-source robotics frame- work for research and benchmarking. arXiv preprint arXiv:1906.08236, 2019

    1906 arXiv

  33. [41]

    Survey: Robot programming by demonstration

    Aude Billard, Sylvain Calinon, Ruediger Dillmann, and Stefan Schaal. Survey: Robot programming by demonstration. Springer handbook of robotics , pages 1371–1394, 2008

    2008

  34. [42]

    Kinesthetic teaching in assembly operations–a user study

    Arne Muxfeldt, Jan-Henrik Kluth, and Daniel Kubus. Kinesthetic teaching in assembly operations–a user study . InSimulation, Modeling, and Programming for Autonomous Robots: 4th International Conference, SIMPAR 2014, Bergamo, Italy, October 20-23,

    2014

  35. [43]

    How to train your robots? the impact of demonstration modality on imitation learning, 2025

    Haozhuo Li, Yuchen Cui, and Dorsa Sadigh. How to train your robots? the impact of demonstration modality on imitation learning, 2025. URL https://arxiv.org/ abs/2503.07017. 28

    2025 arXiv

  36. [44]

    Remote telemanipulation with adapting viewpoints in visually complex environments

    Daniel Rakita, Bilge Mutlu, and Michael Gleicher. Remote telemanipulation with adapting viewpoints in visually complex environments. In Proceedings of Robotics: Science and Systems, FreiburgimBreisgau, Germany, 2019. doi: 10.15607/RSS.2019.XV . 068

    2019 doi

  37. [45]

    Mower, Wolfgang Merkt, Aled Davies, and Sethu Vijayakumar

    Christopher E. Mower, Wolfgang Merkt, Aled Davies, and Sethu Vijayakumar. Com- paring alternate modes of teleoperation for constrained tasks. In2019 IEEE 15th Inter- national Conference on Automation Science and Engineering (CASE), pages 1497–1504,

  38. [46]

    Springer, 2014

    Proceedings 4, pages 533–544. Springer, 2014

    2014

  39. [47]

    Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn

    T ony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware, 2023. URL https://arxiv.org/ abs/2304.13705

    2023 arXiv

  40. [48]

    Helong Huang, Christopher E. Mower, Guowei Huang, Sarthak Das, Magnus Dierk- ing, Guangyuan Luo, Kai T an, Xi Chen, Yehai Yang, Yingbing Chen, Yiming Zeng, Yinchuan Li, Zhanpeng Zhang, Shuang Wu, Yingxue Zhang, Weichao Qiu, T ongtong Cao, Yuzheng Zhuang, Guangjian Tian, Jianye...

    2025

  41. [49]

    Fastslam: An efficient solution to the simultaneous localization and mapping problem with unknown data association

    Sebastian Thrun, Michael Montemerlo, Daphne Koller, Ben Wegbreit, Juan Nieto, and Eduardo Nebot. Fastslam: An efficient solution to the simultaneous localization and mapping problem with unknown data association. Journal of Machine Learning Re- search (JMLR), 2004

    2004

  42. [50]

    Code as policies: Language model programs for embodied con- trol

    Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Flo- rence, and Andy Zeng. Code as policies: Language model programs for embodied con- trol. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493–9500. IEEE, 2023. 29

    2023

  43. [51]

    An optimization-based formalism for shared autonomy in dy- namic environments

    Christopher Mower. An optimization-based formalism for shared autonomy in dy- namic environments. 2022

    2022

  44. [1987]

    doi: 10.1109/JRA.1987.1087068

    1987

  45. [2014]

    URL https://arxiv.org/abs/1410.0759

    arXiv

  46. [2019]

    doi: 10.1109/COASE.2019.8843265

    2019

  47. [2024]

    URL https://arxiv.org/abs/2410.24164

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.