Benchmark¶
RLightning is evaluated on two representative embodied RL tasks covering both small-model high-frequency control and large-model inference settings.
Experimental Setup¶
Hardware
Cluster |
Spec |
Algorithm |
|---|---|---|
H200 cluster |
2 nodes × 8 × H200 141 GB GPUs; |
OpenVLA PPO (large-model manipulation) |
RTX 4090 cluster |
8 nodes × 8 × RTX 4090 GPUs; |
Humanoid WBC (small-model locomotion) |
Software: Ray 2.46.0, PyTorch 2.6, CUDA 12.4, Isaac Lab 2.2.0, mani-skill 3.0.0b21.
Baselines
BeyondMimic — state-of-the-art open-source implementation for humanoid whole-body control. Single-process only; compared at 1 GPU.
RLinf — distributed RL framework; OpenVLA-RL ported to RLightning and compared at 8 GPUs.
OpenVLA PPO¶
Key findings:
RLightning achieves comparable convergence accuracy with benchmark.
RLightning converges to equivalent accuracy approximately ~1.3× faster in wall-clock time compared to RLinf.
Humanoid Whole-Body Control¶
Scalability is measured on the humanoid WBC task across intra-node and inter-node configurations.
In the throughput-intensive humanoid whole-body control task, RLightning maintains throughput efficiency on par with the baseline in the single-GPU setting. Furthermore, with configuration-only changes and no code rewriting, the same training pipeline scales smoothly to 2, 4, and 8 nodes, reaching up to 15× the data throughput of the single-process setup.