Tag: Reinforcement Learning

9 entries

MuJoCo

General-Purpose Physics Engine / Contact Dynamics Solver and Robotics Simulation Backend
An open-source physics engine maintained by Google DeepMind, pairing generalised coordinates with soft convex contact dynamics for robotics, biomechanics and reinforcement learning, with MJCF, MJX, MuJoCo Warp and the 4.0 release.
New Apache-2.0 for source code; CC BY 4.0 for documentation, images and videos in the doc directory Physics EngineContact DynamicsRobot Simulation

Tokamak Plasma Magnetic Confinement Control

Energy / Fusion Control
Deep reinforcement learning controls 19 magnetic coils of the Swiss TCV tokamak at 10 kHz to hold plasma shapes, snowflake divertor and droplet configurations, one of the most complete public sim-to-real control loops on fusion hardware.
New Paper copyright retained by the authors; algorithm and libraries under their own open-source licenses FusionReinforcement LearningPlasma Control

HALO

Humanoid Robot Autonomy System / Sim-to-Real Motion Generation
Apple's open-code perceptive humanoid autonomy system: trained entirely in simulation and transferred to a Unitree G1, achieving 91% single-task success in unseen real homes.
New Apple proprietary license (non-OSI; GitHub marks Other/NOASSERTION; see repository LICENSE) Embodied AIHumanoid RobotSim-to-Real

Steel

GPU Rigid-Body Dynamics Solver / Heterogeneous-Batch Backend for the Newton Physics Engine
A GPU rigid-body dynamics solver open sourced by Skild AI with NVIDIA and contributed to the Newton physics engine, built for batches that mix different robots, tasks and worlds; the project reports 3.8x PhysX geometric-mean simulation throughput across 20 benchmarks.
New Apache-2.0 for Steel and Newton source code; CC BY 4.0 for Newton documentation and separate terms for bundled assets Rigid-Body DynamicsPhysics EngineRobot Simulation

SuperMimic

Humanoid Whole-Body Control / Differentiable World Model Training Framework
An open-source humanoid whole-body control framework from UC Berkeley that amortizes learning across all motions with one differentiable world model, surpassing a 128-GPU, 7-day baseline in 3 hours on a single GPU and winning the CoRL 2026 Best Paper Award.
New Apache License 2.0 Physics SimulationHumanoid RobotMotion Imitation

NVIDIA ProtoMotions

GPU-Accelerated Humanoid / Digital-Human Motion Simulation and Learning Framework
An open-source GPU-accelerated simulation and learning framework from NVIDIA Labs for digital humans and humanoid robots, covering motion tracking, generative policies and sim-to-real deployment across IsaacGym, Isaac Lab, Newton and MuJoCo backends.
New Apache License 2.0 Physics SimulationHumanoid RobotMotion Generation

HydroGym: A Reinforcement Learning Benchmark Platform for Flow Control

Flow Control / Active Flow Control
A solver-independent reinforcement learning platform for flow control, providing 61 validated open-source 2D and 3D environments across six CFD solver backends including fully differentiable solvers. Its zero-shot transfer experiment deploys a channel-trained policy onto a 3D NACA0012 wing, cutting local skin friction by about 38% and total drag by about 11% while cutting exploration cost by four orders of magnitude relative to direct on-wing training.
New MIT (code) / Nature paper Flow ControlReinforcement LearningDrag Reduction

Wind Turbine Aeroelastic Simulation and Wake Control: A Verifiable Chain from OpenFAST to Reinforcement Learning

Wind Energy / Multiphysics Simulation / Wake Control
A verifiable chain through OpenFAST aeroelastic simulation, the FAST.Farm multi-turbine wake model, the FLORIS engineering wake model and IEA Wind reference machines. This entry checks the origin of every quantitative claim and records where the chain breaks.
New Apache-2.0 for OpenFAST, FAST.Farm, FLORIS and the IEA Wind Task 55 reference models Wind EnergyAeroelasticityWake Control

Data-Driven MPC with Adaptively Sampled Non-Expert Demonstrations: Performance Guarantees and Sample Complexity

Multibody
This paper proposes a data-driven policy construction method for dynamical systems using non-expert MPC demonstrations, introducing a memory-based nonparametric policy with Lipschitz-regularized upper envelope of the finite-horizon surrogate value function to approximate infinite-horizon optimal control.
See LICENSE file Model Predictive ControlData-Driven ControlReinforcement Learning

← All Tags