Scenes and robotics: paper-to-code routes
On this page
All 60 tools · How to contribute
These are construction routes researched by agents. The execution line reports the bounded work actually performed; the comparison section defines what remains before claiming a published result. Human scientific review is pending.
openusd#
OpenUSD — Can a paper-associated robot scene preserve the model and time-varying experiment metadata through USD?
Route: Paper model adaptation. Execution: Not run. Source confidence: medium.
Paper for openusd#
Mayank Mittal et al. (2023). Orbit: A Unified Simulation Framework for Interactive Robot Learning Environments. 10.1109/LRA.2023.3270034.
Publication type: journal_article. Primary metadata. Metadata verification: confirmed. Full-text access: Live Crossref metadata refreshed; prior lab inspected open manuscript architecture
Target in the paper: USD scene architecture; current successor Cartpole USD asset is an adaptation, not the 2023 experiment archive.
Code and inputs for openusd#
Artifact or upstream source. Isaac Lab is Orbit successor; inspected source loads Cartpole/cartpole.usd through UsdFileCfg. USD supports scene exchange; PhysX solves dynamics.
Source inspected: yes. Ref: v2.2.0. Commit: 46dff135f44683f031edf346e544fcfd8456b2bb.
Files: source/isaaclab_assets/isaaclab_assets/robots/cartpole.py; source/isaaclab_tasks/isaaclab_tasks/direct/cartpole/cartpole_env.py
Terms: BSD-3-Clause source; separate NVIDIA runtime and asset terms
Environment and compute: Isaac Lab v2.2.0 with compatible Isaac Sim 5.0.0; standalone usd-core for inspection CPU for USD inspection; NVIDIA GPU for coupled Isaac simulation.
Inputs: Cartpole USD asset referenced by upstream configuration; preserve resolved dependency hashes and simulator settings.
Data access: Source public; robot assets and runtime carry separate terms.
Construct and compare openusd#
The sequence below contains upstream commands and proposed setup instructions, as labeled. Only commands identified in the execution evidence were run here; inspect the pinned source and prerequisites before running.
Proposed: fetch the configured Cartpole USD asset through the documented Isaac asset endpoint and retain a local dependency manifest.Proposed: use pxr.Usd.Stage.Open(local_asset), inspect articulation/physics schemas and reference dependencies, and compare attributes before/after export.
Expected outputs: Editable USD scene, dependency manifest, source-to-export attribute comparison.
Comparison with the paper: Verify composed prim paths, units, joint names and physics attributes are preserved. No paper numerical outcome is claimed.
Execution result: No execution claim; commands were inspected upstream or explicitly marked proposed.
Execution evidence: None recorded.
Blockers: Original Orbit experiment version not matched; local USD asset retrieval not attempted.
Next action: Develop the scene-manifest exercise beside the existing battery-results-to-USD example.
Gaps: Original Orbit experiment version not matched; local USD asset retrieval not attempted.
Sources for openusd#
- Paper identity (inspected 2026-11-15).
- Concrete published-framework USD scene loading configuration (inspected 2026-11-15).
blender#
Blender — Can we regenerate MOVi-A synthetic videos and their ground-truth annotations?
Route: Paper result candidate. Execution: Example executed. Source confidence: high.
Paper for blender#
Klaus Greff et al. (2022). Kubric: A scalable dataset generator. arXiv:2203.03570v1.
Publication type: conference_paper. Primary metadata. Metadata verification: confirmed. Full-text access: Open CVPR 2022 manuscript HTML inspected
Target in the paper: Section 4.1 and Figure 5, MOVi-A scene-generation properties; Table 3 learning results require a separate training reproduction.
Code and inputs for blender#
Artifact or upstream source. Authors explicitly identify MOVi as a dataset contribution of their CVPR paper; Blender renders and PyBullet simulates.
Source inspected: yes. Ref: main at inspection. Commit: 61f2422c84bab75006df33c6989e0b483db3ccfe.
Files: challenges/movi/README.md; challenges/movi/movi_ab_worker.py; challenges/movi/movi_a.py
Terms: Kubric Apache-2.0; Blender GPL; inspect bundled object/material terms
Environment and compute: Authors kubricdockerhub/kubruntu image; upstream documents Blender 2.93. Resolve/pin image digest before production reproduction. CPU rendering possible for one scene; full dataset rendering is expensive.
Inputs: CLEVR-style object set, background and camera; one fixed scene seed initially, then matching dataset generator settings.
Data access: Public authors code and public dataset bucket; retain asset manifest and terms.
Construct and compare blender#
The sequence below contains upstream commands and proposed setup instructions, as labeled. Only commands identified in the execution evidence were run here; inspect the pinned source and prerequisites before running.
Upstream README: git clone https://github.com/google-research/kubric.gitProposed pin: git checkout 61f2422c84bab75006df33c6989e0b483db3ccfeUpstream MOVi-A command: docker run --rm --interactive --user $(id -u):$(id -g) --volume "$(pwd):/kubric" kubricdockerhub/kubruntu /usr/bin/python3 challenges/movi/movi_ab_worker.py --objects_set=clevr --background=clevr --camera=clevr
Expected outputs: RGB frames, instance segmentation, depth, optical flow, scene metadata and saved scene as configured by worker.
Comparison with the paper: PublishedMOVi-A dataset-design properties match; all9numerical/design checks passed. Originalfigurepixels, originalassetbytes andTable3learning scores are outside this comparison.
Execution result: Paper-date author worker ran to completion onCPU:24frames,256×256,7objects and all rendered annotations/native states. Nine design checks passed; second scene has identical instance metadata and24/24RGBAhashes.
Execution evidence: Author-code attempt; Run record
Blockers: Exact2022assetbytes and full originaldependencylock not established; SAVi/SIMONeTable3training benchmark not attempted.
Next action: Human review of the dataset-generation guide; match archivedassets and modeltraining protocol before reproducingTable3.
Gaps: Exact2022assetbytes and full originaldependencylock not established; SAVi/SIMONeTable3training benchmark not attempted.
Sources for blender#
- Paper identity (inspected 2026-11-15).
- Section 4.1 and Figure 5 (inspected 2026-11-15).
- Author generator command and dataset properties (inspected 2026-11-15).
isaacsim#
Isaac Sim — Can the published framework train a physics-based classic-control policy?
Route: Software paper example. Execution: Not run. Source confidence: medium.
Paper for isaacsim#
NVIDIA; Mayank Mittal et al. (2025). Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning. arXiv:2511.04831v1.
Publication type: preprint. Primary metadata. Metadata verification: confirmed. Full-text access: Open preprint HTML inspected; publication beyond preprint not established
Target in the paper: Section 3.7.3 classic Cartpole environments; Section 4.1 throughput definition. This route does not reproduce hardware throughput figures.
Code and inputs for isaacsim#
Artifact or upstream source. Paper directly links this author framework. Inspected v2.2.0 is a concrete supported baseline; correspondence to the paper benchmark commit is unestablished.
Source inspected: yes. Ref: v2.2.0. Commit: 46dff135f44683f031edf346e544fcfd8456b2bb.
Files: source/isaaclab_tasks/isaaclab_tasks/direct/cartpole/cartpole_env.py; source/isaaclab_tasks/isaaclab_tasks/direct/cartpole/agents/rsl_rl_ppo_cfg.py; scripts/reinforcement_learning/rsl_rl/train.py; docs/source/overview/reinforcement-learning/rl_existing_scripts.rst
Terms: Isaac Lab BSD-3-Clause core; Isaac Sim source Apache-2.0; runtime/assets separate terms
Environment and compute: Isaac Lab v2.2.0; Isaac Sim 5.0.0 and supported NVIDIA GPU/driver; RSL-RL wrapper. NVIDIA GPU runtime; no hardware-scale training attempted.
Inputs: Isaac-Cartpole-Direct-v0, explicit seed and recorded number of environments; task dt=1/120 s and decimation=2 in inspected config.
Data access: Public source; network robot assets and NVIDIA runtime terms.
Construct and compare isaacsim#
The sequence below contains upstream commands and proposed setup instructions, as labeled. Only commands identified in the execution evidence were run here; inspect the pinned source and prerequisites before running.
Proposed pin: git checkout 46dff135f44683f031edf346e544fcfd8456b2bbUpstream wrapper installation route: ./isaaclab.sh -i rsl_rlProposed task selection using inspected train entry point: ./isaaclab.sh -p scripts/reinforcement_learning/rsl_rl/train.py --task Isaac-Cartpole-Direct-v0 --num_envs 64 --seed 42 --headless
Expected outputs: Training configuration, logs and policy checkpoints; deterministic evaluation episodes; optional USD scene manifest.
Comparison with the paper: Compare multiple same-seed and different-seed reward/episode-length curves. A 64-environment teaching run is not the paper throughput experiment; matching Section 4.1 requires exact task, GPU, environment count and timing definition.
Execution result: No execution claim; commands were inspected upstream or explicitly marked proposed.
Execution evidence: None recorded.
Blockers: No local compatible GPU runtime verified; paper benchmark revision/hardware not matched.
Next action: Provision an approved GPU runner, pin the paper environment, then run a small complete training/evaluation cycle.
Gaps: No local compatible GPU runtime verified; paper benchmark revision/hardware not matched.
Sources for isaacsim#
- Paper metadata and direct author-code association (inspected 2026-11-15).
- Section 3.7.3 names classic tasks; Section 4.1 defines throughput (inspected 2026-11-15).
- Concrete model timestep/rewards/reset configuration (inspected 2026-11-15).
- Training wrapper route (inspected 2026-11-15).
isaaclab#
Isaac Lab — Can we train, save and evaluate a paper-associated Cartpole policy reproducibly?
Route: Software paper example. Execution: Not run. Source confidence: medium.
Paper for isaaclab#
NVIDIA; Mayank Mittal et al. (2025). Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning. arXiv:2511.04831v1.
Publication type: preprint. Primary metadata. Metadata verification: confirmed. Full-text access: Open preprint HTML inspected; publication beyond preprint not established
Target in the paper: Section 3.7.3 classic Cartpole environments; Section 4.1 throughput definition. This route does not reproduce hardware throughput figures.
Code and inputs for isaaclab#
Artifact or upstream source. Paper directly links this author framework. Inspected v2.2.0 is a concrete supported baseline; correspondence to the paper benchmark commit is unestablished.
Source inspected: yes. Ref: v2.2.0. Commit: 46dff135f44683f031edf346e544fcfd8456b2bb.
Files: source/isaaclab_tasks/isaaclab_tasks/direct/cartpole/cartpole_env.py; source/isaaclab_tasks/isaaclab_tasks/direct/cartpole/agents/rsl_rl_ppo_cfg.py; scripts/reinforcement_learning/rsl_rl/train.py; docs/source/overview/reinforcement-learning/rl_existing_scripts.rst
Terms: Isaac Lab BSD-3-Clause core; Isaac Sim source Apache-2.0; runtime/assets separate terms
Environment and compute: Isaac Lab v2.2.0; Isaac Sim 5.0.0 and supported NVIDIA GPU/driver; RSL-RL wrapper. NVIDIA GPU runtime; no hardware-scale training attempted.
Inputs: Isaac-Cartpole-Direct-v0, explicit seed and recorded number of environments; task dt=1/120 s and decimation=2 in inspected config.
Data access: Public source; network robot assets and NVIDIA runtime terms.
Construct and compare isaaclab#
The sequence below contains upstream commands and proposed setup instructions, as labeled. Only commands identified in the execution evidence were run here; inspect the pinned source and prerequisites before running.
Proposed pin: git checkout 46dff135f44683f031edf346e544fcfd8456b2bbUpstream wrapper installation route: ./isaaclab.sh -i rsl_rlProposed task selection using inspected train entry point: ./isaaclab.sh -p scripts/reinforcement_learning/rsl_rl/train.py --task Isaac-Cartpole-Direct-v0 --num_envs 64 --seed 42 --headless
Expected outputs: Training configuration, logs and policy checkpoints; deterministic evaluation episodes; optional USD scene manifest.
Comparison with the paper: Compare multiple same-seed and different-seed reward/episode-length curves. A 64-environment teaching run is not the paper throughput experiment; matching Section 4.1 requires exact task, GPU, environment count and timing definition.
Execution result: No execution claim; commands were inspected upstream or explicitly marked proposed.
Execution evidence: None recorded.
Blockers: No local compatible GPU runtime verified; paper benchmark revision/hardware not matched.
Next action: Provision an approved GPU runner, pin the paper environment, then run a small complete training/evaluation cycle.
Gaps: No local compatible GPU runtime verified; paper benchmark revision/hardware not matched.
Sources for isaaclab#
- Paper metadata and direct author-code association (inspected 2026-11-15).
- Section 3.7.3 names classic tasks; Section 4.1 defines throughput (inspected 2026-11-15).
- Concrete model timestep/rewards/reset configuration (inspected 2026-11-15).
- Training wrapper route (inspected 2026-11-15).
mujoco#
MuJoCo — Can we run the paper Cartpole environment and reproduce its explicit quaternion example?
Route: Software paper example. Execution: Example executed. Source confidence: high.
Paper for mujoco#
Yuval Tassa et al. (2018). DeepMind Control Suite. arXiv:1801.00690v1.
Publication type: preprint. Primary metadata. Metadata verification: confirmed. Full-text access: Open manuscript HTML inspected
Target in the paper: Section 4 suite.load Cartpole swingup and reset/step episode example; Section 5 quaternion-to-matrix example.
Code and inputs for mujoco#
Artifact or upstream source. Paper explicitly releases DeepMind Control Suite, powered by MuJoCo. Modern package differs from 2018 engine.
Source inspected: yes. Ref: main at inspection. Commit: d7e52e3602a555df6cc20cc928ea5b41e9920158.
Files: dm_control/suite/cartpole.py; dm_control/suite/cartpole.xml; README.md
Terms: Apache-2.0; MuJoCo Apache-2.0
Environment and compute: Isolated Python 3.12; dm-control==1.0.52; freeze all resolved MuJoCo dependencies. CPU; omit rendering to avoid display dependency.
Inputs: Paper Cartpole swingup; seed=42 for environment and action generator; quaternion (0.5,0.5,0.5,0.5).
Data access: Public code and bundled MJCF assets.
Construct and compare mujoco#
The sequence below contains upstream commands and proposed setup instructions, as labeled. Only commands identified in the execution evidence were run here; inspect the pinned source and prerequisites before running.
Upstream installation: python -m pip install dm-control==1.0.52Paper API route: from dm_control import suite; env = suite.load(domain_name="cartpole", task_name="swingup"); env.reset(); env.step(action)Proposed modern binding for paper Section 5: mujoco.mju_quat2Mat(matrix, quaternion)
Expected outputs: One full episode trajectory and summary; quaternion matrix compared with explicitly printed paper matrix.
Comparison with the paper: Paper matrix is [[0,0,1],[1,0,0],[0,1,0]]. Check exactly with tolerance 1e-12; run the paper episode loop and verify finite states/rewards. These checks do not reproduce learning-algorithm performance curves.
Execution result: Modern dm-control 1.0.52: 1000 Cartpole swingup steps over 9.999999999999831 simulated seconds; two seeded episodes exactly identical. Paper Section 5 quaternion matrix matched with zero error (chosen absolute tolerance 1e-12). This is execution of paper software examples, not learning-benchmark reproduction.
Execution evidence: Run receipt; Numeric output; Numeric output
Blockers: Modern engine/package differs from 2018; original learning benchmark not attempted.
Next action: Execute and retain full episode CSV, package lock, matrix comparison, source and output hashes.
Gaps: Modern engine/package differs from 2018; original learning benchmark not attempted.
Sources for mujoco#
- Paper metadata (inspected 2026-11-15).
- Actual example API and printed matrix (inspected 2026-11-15).
- Author task implementation (inspected 2026-11-15).
- Installation and supported simulation backend (inspected 2026-11-15).
gz-sim#
Gazebo — Do isolated same-seed PPO runs produce identical checkpoints and evaluation curves?
Route: Paper result candidate. Execution: Example executed. Source confidence: high.
Paper for gz-sim#
Amal Dev Haridevan; Junjie Kang; Jinjun Shan (2026). GzDRL: Reproducible and Scalable Deep Reinforcement Learning with Gazebo. arXiv:2609.13243v1.
Publication type: preprint. Primary metadata. Metadata verification: confirmed. Full-text access: Open preprint HTML inspected; manuscript says submitted to IEEE, publication not established
Target in the paper: Section V-B, Figure 6, Table II: process-isolated same-seed and different-seed PPO reproducibility.
Code and inputs for gz-sim#
Artifact or upstream source. Exact author paper artifact, explicitly labeled research artifact; paper links same repository.
Source inspected: yes. Ref: HEAD at inspection. Commit: 3e5c5f9a62c1160b3b2889e6d34d794a5361e178.
Files: experiments_paper/reproducability_validation/README.md; experiments_paper/reproducability_validation/config.json; experiments_paper/reproducability_validation/run_experiment.py; experiments_paper/reproducability_validation/verify_determinism.py
Terms: MIT; Gazebo/dependency notices separate
Environment and compute: Upstream Gazebo/DRLServer C++ build prerequisites plus Python RL extra; freeze compiler, Gazebo and SB3 versions. CPU physics and PPO; 10 full training runs. Paper hardware must be recorded for throughput claims.
Inputs: Author config: 5 same-seed runs with seed 42; different seeds {1,5,25,42,125}; eight environments; 500000 timesteps; checkpoints every 50000; 20 deterministic evaluation episodes.
Data access: Public code and bundled config; no private data indicated.
Construct and compare gz-sim#
The sequence below contains upstream commands and proposed setup instructions, as labeled. Only commands identified in the execution evidence were run here; inspect the pinned source and prerequisites before running.
Upstream installation, from repository root after native build: python -m pip install -e '.[rl]'cd experiments_paper/reproducability_validationpython run_experiment.py
Expected outputs: results/determinism_check.json; per-run checkpoint/evaluation data; analysis CSV, plots, LaTeX table and machine/config manifest.
Comparison with the paper: Same-seed determinism is supported by full author-code execution and independent parameter hashing. The published numerical final return is not reproduced; no tolerance is invented. Historical package/compiler lock is not established, and this portablebuild uses Gazebo10.6/GCC15 plus a mechanical constructor-disambiguation patch. Different-seed row was not run. Human scientific review pending.
Execution result: Completed all5full same-seed500000-requested-step training runs and1200evaluation episodes. Author preflight and independent12-checkpoint parameter identity checks pass; finalreturn16.53217 differs from paper18.43.
Execution evidence: Author-code attempt; Run record
Blockers: Paper final-return18.43not reproduced (actual16.53217); original package/compiler lock and exactpaperartifactversionnot established.; Different-seedTableIIrownot run.
Next action: Review native task/config/environment differences with paper authors; reproduce the original score only after exact software/config provenance is resolved.
Gaps: Paper final-return18.43not reproduced (actual16.53217); original package/compiler lock and exactpaperartifactversionnot established.; Different-seedTableIIrownot run.
Sources for gz-sim#
- Paper metadata (inspected 2026-11-15).
- Specific Table II result and full experiment settings (inspected 2026-11-15).
- Exact author run commands and output contract (inspected 2026-11-15).
webots#
Webots — Can a Webots Cartpole learn balancing through the paper Deepbots interface?
Route: Paper model adaptation. Execution: Not run. Source confidence: medium.
Paper for webots#
Manos Kirtas; Konstantinos Tsampazis; Nikolaos Passalis; Anastasios Tefas (2020). Deepbots: A Webots-Based Deep Reinforcement Learning Framework for Robotics. 10.1007/978-3-030-49186-4_6.
Publication type: conference_paper. Primary metadata. Metadata verification: confirmed. Full-text access: PMC article search text and author docs inspected; direct PMC open was access challenged
Target in the paper: Section 4 use cases and Cartpole code examples; exact published learning-curve settings require full-text comparison.
Code and inputs for webots#
Artifact or upstream source. Official Deepbots documentation links deepworlds as its example repository; current examples implement author framework but are not a pinned 2020 artifact.
Source inspected: yes. Ref: HEAD at inspection. Commit: 4f21e555216750a4aafb84baad261aeda7cb1f74.
Files: examples/cartpole/README.md; examples/cartpole/cartpole_discrete/worlds/cartpole_world_robot_supervisor.wbt; examples/cartpole/cartpole_discrete/controllers/robot_supervisor_manager/PPO_runner.py
Terms: Review deepworlds/deepbots license and Webots asset notices before copying code; Webots Apache-2.0
Environment and compute: Webots plus matching Python deepbots/PyTorch/Gym requirements from chosen example. CPU simulation and small neural network; Webots native runtime required.
Inputs: Discrete-action Cartpole world, robot-supervisor scheme and PPO agent; retain seeds, episode lengths and reset settings.
Data access: Public author code and robot world; no private dataset.
Construct and compare webots#
The sequence below contains upstream commands and proposed setup instructions, as labeled. Only commands identified in the execution evidence were run here; inspect the pinned source and prerequisites before running.
Upstream concrete route: install the selected example requirements, open examples/cartpole/cartpole_discrete/worlds/cartpole_world_robot_supervisor.wbt in Webots, and run the supervisor PPO controller.Proposed automation: record environment/controller versions, log episode reward and evaluate a frozen checkpoint over fixed-seed episodes.
Expected outputs: Episode reward plot and trained agent as illustrated in authors README; retain policy and per-episode logs.
Comparison with the paper: Compare against paper Section 4 Cartpole result after recovering its original settings; current README rewardPlot.png is a lead, not an established paper numeric target.
Execution result: No execution claim; commands were inspected upstream or explicitly marked proposed.
Execution evidence: None recorded.
Blockers: Historical Deepbots/Webots/Gym versions not matched; paper curve not digitized; native runtime unavailable.
Next action: Retrieve full text via legitimate route and map the current world to the paper use case before accepting a reproduction target.
Gaps: Historical Deepbots/Webots/Gym versions not matched; paper curve not digitized; native runtime unavailable.
Sources for webots#
- Title/authors/year/DOI verified live (inspected 2026-11-15).
- Author framework citation and official deepworlds link (inspected 2026-11-15).
- Concrete world/controller routes and existing reward plot (inspected 2026-11-15).
bullet3#
Bullet / PyBullet — Can the released Minitaur trotting and galloping controllers be replayed?
Route: Paper result candidate. Execution: Example executed. Source confidence: high.
Paper for bullet3#
Jie Tan; Tingnan Zhang; Erwin Coumans; Atil Iscen; Yunfei Bai; Danijar Hafner; Steven Bohez; Vincent Vanhoucke (2018). Sim-to-Real: Learning Agile Locomotion For Quadruped Robots. arXiv:1804.10332v2.
Publication type: conference_paper. Primary metadata. Metadata verification: confirmed. Full-text access: Open abstract and author environment README inspected; detailed experimental section not audited
Target in the paper: Paper trotting and galloping policies; authors README explicitly identifies the two exact environments used.
Code and inputs for bullet3#
Artifact or upstream source. Author README explicitly states these two environments were used in the paper and links pre-trained PPO examples.
Source inspected: yes. Ref: master at inspection. Commit: d3afe76132fc76b0f9b2359cbaef47a3766e7fbb.
Files: examples/pybullet/gym/pybullet_envs/minitaur/envs/README.md; examples/pybullet/gym/pybullet_envs/minitaur/envs/minitaur_reactive_env.py; examples/pybullet/gym/pybullet_envs/minitaur/envs/minitaur_trotting_env.py; examples/pybullet/gym/pybullet_data/policies/ppo/minitaur_trotting_env/config.yaml
Terms: Zlib core; separate example/model/checkpoint dependency notices
Environment and compute: Historical TensorFlow-1-style PPO environment, Gym, ruamel.yaml and compatible PyBullet; inspect imports and resolve original versions before installation. CPU policy replay possible with legacy-compatible dependencies; training and physical testing are separate.
Inputs: Bundled Minitaur URDF, PPO checkpoint/config and paper motor/latency/randomization configuration.
Data access: Public source, bundled checkpoints and model; review individual asset notices.
Construct and compare bullet3#
The sequence below contains upstream commands and proposed setup instructions, as labeled. Only commands identified in the execution evidence were run here; inspect the pinned source and prerequisites before running.
Upstream README: pip install gym; pip install ruamel.yamlUpstream, from example directory with dependencies available: python minitaur_reactive_env_example.pyUpstream: python minitaur_trotting_env_example.py
Expected outputs: Policy rollouts, robot states/actions and gait trajectories; archive raw logs and exact checkpoint hashes.
Comparison with the paper: Released policies execute, but published simulation speeds are not reproduced by matched methods/checkpoints. Trot mean0.9215m/s versus paper approximately0.50; gallop mean1.3084 with SD0.5209 and only2/5 complete. No physical sim-to-real claim.
Execution result: Replayed author frozen PPO policies for five seeds per gait; all numerical traces retained.
Execution evidence: Author-code attempt; Run record
Blockers: Published TableII policy widths/checkpoint selection not matched; native physics revision not2018; no physical robot evaluation.
Next action: Recover paper-specific checkpoint/environment; agree the speed estimator and tolerance, then compare with repeated episodes.
Gaps: Published TableII policy widths/checkpoint selection not matched; native physics revision not2018; no physical robot evaluation.
Sources for bullet3#
- Primary metadata and declared trotting/galloping outcomes (inspected 2026-11-15).
- Explicit author association, released environments and replay commands (inspected 2026-11-15).
chrono#
Project Chrono — How do steering and terrain inputs affect a template-based HMMWV model?
Route: Software paper example. Execution: Not run. Source confidence: medium.
Paper for chrono#
Radu Serban; Michael Taylor; Dan Negrut; Alessandro Tasora (2019). Chrono::Vehicle: template-based ground vehicle modelling and simulation. 10.1504/IJVP.2019.097096.
Publication type: journal_article. Primary metadata. Metadata verification: confirmed. Full-text access: Publisher abstract and metadata inspected; full text not inspected
Target in the paper: Software architecture and ground-vehicle modeling demonstration; exact paper figure/model identification pending full text.
Code and inputs for chrono#
Artifact or upstream source. Official current HMMWV example is by Chrono authors and uses Chrono::Vehicle, the module described in the paper. It is not established as the exact 2019 experiment.
Source inspected: yes. Ref: main at inspection. Commit: c2865ec31e5913838898bf9c13b1bba66804db3e.
Files: src/demos/python/vehicle/demo_VEH_HMMWV.py
Terms: BSD-3-Clause code; model/data notices separate
Environment and compute: PyChrono with vehicle and Irrlicht modules, matching supplied vehicle data; prefer a supported binary environment over rebuilding all modules. CPU multibody dynamics; display support for unmodified Irrlicht demo.
Inputs: Upstream HMMWV_Full, terrain, driver and tire settings in demo; replace interactive driver with a saved deterministic input only as a documented adaptation.
Data access: Public model code and bundled vehicle data.
Construct and compare chrono#
The sequence below contains upstream commands and proposed setup instructions, as labeled. Only commands identified in the execution evidence were run here; inspect the pinned source and prerequisites before running.
Proposed from a matching PyChrono environment: python src/demos/python/vehicle/demo_VEH_HMMWV.py
Expected outputs: Interactive vehicle demo; for reproducibility add logged vehicle position, speed, steering input and model configuration.
Comparison with the paper: First check finite state and repeatability under recorded driver input. Identify the paper figure and mesh/tire/terrain settings before comparing published trajectories or response curves.
Execution result: No execution claim; commands were inspected upstream or explicitly marked proposed.
Execution evidence: None recorded.
Blockers: Full article and exact published model/response not inspected; demo is interactive and current API differs from 2019.
Next action: Obtain author manuscript from official Chrono whitepaper list, extract a target figure, then choose its matching model version.
Gaps: Full article and exact published model/response not inspected; demo is interactive and current API differs from 2019.
Sources for chrono#
- Verified live DOI/title/authors (inspected 2026-11-15).
- Primary publication record and abstract (inspected 2026-11-15).
- Concrete author HMMWV model and dependencies (inspected 2026-11-15).
sofa#
SOFA — How does tendon displacement bend a finite-element soft finger?
Route: Paper model adaptation. Execution: Not run. Source confidence: medium.
Paper for sofa#
Eulalie Coevoet et al. (2017). Software toolkit for modeling, simulation, and control of soft robots. 10.1080/01691864.2017.1395362.
Publication type: journal_article. Primary metadata. Metadata verification: confirmed. Full-text access: Publisher abstract and author publication metadata inspected; full model/results not inspected
Target in the paper: Direct forward soft-robot simulation described in article; current CableConstraint/Finger.py is an author-associated teaching model, not a verified exact figure.
Code and inputs for sofa#
Artifact or upstream source. The paper presents the SOFA-based toolkit. Current official plugin source supplies a cable-driven FEM finger; paper-specific historical mesh and constants are unverified.
Source inspected: yes. Ref: master at inspection. Commit: 3dadbaff9f50b379bf1cc680b379b74810751ee0.
Files: examples/component/constraint/CableConstraint/Finger.py; examples/component/constraint/CableConstraint/FingerController.py; examples/component/constraint/CableConstraint/mesh/finger.vtk
Terms: SoftRobots LGPL-3.0 according to live repository license metadata; SOFA LGPL-2.1; dependencies separate
Environment and compute: SOFA with matching SoftRobots and SofaPython3 plugins; preserve full component/plugin versions. CPU FEM; SOFA native runtime and plugins; optional display.
Inputs: Bundled finger mesh; source sets dt=0.01 and gravity=[0,-9810,0] in model units; inspect and retain material, constraint and actuator definitions.
Data access: Public source and supplied mesh; verify mesh redistribution terms.
Construct and compare sofa#
The sequence below contains upstream commands and proposed setup instructions, as labeled. Only commands identified in the execution evidence were run here; inspect the pinned source and prerequisites before running.
Proposed supported-runner route: runSofa examples/component/constraint/CableConstraint/Finger.pyProposed reproducibility adaptation: drive a saved tendon-displacement series and export tip displacement and solver residuals.
Expected outputs: Deformed finger scene and, after instrumentation, displacement-vs-actuation series and convergence logs.
Comparison with the paper: Compare paper actuator/mesh/material settings and a specific deflection curve after full text is obtained. A plausible bending animation or converged linear solve does not establish paper reproduction.
Execution result: No execution claim; commands were inspected upstream or explicitly marked proposed.
Execution evidence: None recorded.
Blockers: Exact 2017 model, mesh, material parameters and numeric target not matched; native plugin stack unavailable.
Next action: Retrieve legitimate author manuscript, map one forward cable-actuation experiment, then instrument exact mesh and loading.
Gaps: Exact 2017 model, mesh, material parameters and numeric target not matched; native plugin stack unavailable.
Sources for sofa#
- Verified live primary citation metadata (inspected 2026-11-15).
- Paper links toolkit and states SOFA modeling relation (inspected 2026-11-15).
- Actual mesh/plugin/actuator scene source (inspected 2026-11-15).