CARL
A CUDA-based GPU-parallel simulation environment for Rocket League that keeps all simulation state in GPU memory, eliminating host-device data transfer overhead.
Timeline (last 90 days)
The applicable conditions for this approach include: the physics logic must be simple enough, it must be efficiently parallelizable, and the task itself must be suited to large-scale environment sampling.
From the training curves, the robot shows a clear improvement in mid-air control capability.
The developer joked that if he were operating it himself, perhaps a single subtle input could have saved it, implying that although the robot's performance is quite good, it is still some distance from a human expert.
CARL is currently "feature complete", and the developer decided to open-source it to the community to see the feedback.
The developer's next TODO list includes: a built-in training package API, speed and physics consistency benchmarks, additional tests and other engineering best practices, and exploring JAX and TensorFlow integration.
The evaluation of physics consistency (parity) is still fairly preliminary, and a more systematic comparison has not yet been completed.
After CARL was posted on Reddit, it attracted a fair amount of attention.
CARL's core idea is to keep all simulation state information in GPU memory, completely eliminating host-to-device data transfer overhead during training.
CARL was inspired by PureJaxRL, which had implemented a similar massively parallel environment scheme on JAX.
CARL is not intended to replace RocketSim, but rather to explore another path of running physics logic in parallel on the GPU using CUDA.
6 more timeline events
All Facts (16)
The agents trained by CARL can be seamlessly transferred to the real game, indicating that its physics simulation maintains good consistency with the actual game.
60%UnverifiedCARL is a GPU-parallel reinforcement learning simulation project that ports the physics logic of Rocket League to CUDA, with the full name CUDA-Accelerated Rocket League.
60%UnverifiedThe applicable conditions for this approach include: the physics logic must be simple enough, it must be efficiently parallelizable, and the task itself must be suited to large-scale environment sampling.
50%UnverifiedCARL was inspired by PureJaxRL, which had implemented a similar massively parallel environment scheme on JAX.
50%UnverifiedCARL is not intended to replace RocketSim, but rather to explore another path of running physics logic in parallel on the GPU using CUDA.
50%UnverifiedFrom the training curves, the robot shows a clear improvement in mid-air control capability.
50%UnverifiedThe developer joked that if he were operating it himself, perhaps a single subtle input could have saved it, implying that although the robot's performance is quite good, it is still some distance from a human expert.
50%UnverifiedThe developer's next TODO list includes: a built-in training package API, speed and physics consistency benchmarks, additional tests and other engineering best practices, and exploring JAX and TensorFlow integration.
50%UnverifiedThe evaluation of physics consistency (parity) is still fairly preliminary, and a more systematic comparison has not yet been completed.
50%UnverifiedCARL's core idea is to keep all simulation state information in GPU memory, completely eliminating host-to-device data transfer overhead during training.
50%UnverifiedThe work still to be advanced in the CARL project includes: a built-in training package API, speed and consistency benchmarks, testing and engineering best practices, and JAX and TensorFlow integration.
50%UnverifiedThe developer said that CARL has now reached a feature-complete state.
50%UnverifiedCARL uses the same training configuration as Nexto (including reward function, hyperparameters and network structure) and reached Nexto's level within hours.
50%UnverifiedAfter CARL was posted on Reddit, it attracted a fair amount of attention.
50%UnverifiedCARL is currently "feature complete", and the developer decided to open-source it to the community to see the feedback.
50%UnverifiedThe CARL project attracted attention after being posted on Reddit.
50%