Humanoid

Humanoid#

Viewer Performance Impact Benchmark

Abstract#

Viewer Performance Impact Benchmark This benchmark quantifies the performance impact of BraxViewer on Brax PPO training speed. We measure training throughput under three scenarios: pure training, viewer backend with rendering disabled, and full real-time visualization.

Methods#

Experimental Design#

We conduct a factorial experiment with three factors:

  • Environment Scale: 64, 256, and 1024 parallel environments

  • Viewer Configuration: No viewer, viewer disabled, and viewer enabled

  • Hardware: Multiple compute devices are tested

Training Settings#

All experiments use identical PPO hyperparameters:

  • Environment: humanoid with positional backend

  • Training steps: 50,000

  • Episode length: 1,000

  • Learning rate: 3e-4

  • Entropy cost: 1e-3

Performance Metrics#

Primary metrics include:

  • Total training time (seconds) (Figure 1, top)

  • Training progress curves (steps vs. time) (Figure 1, bottom)

Results#

Consolidated Benchmark Comparison

Figure 1: Total training time comparison (top) and training progress curves (bottom) across different hardware configurations, environment scales, and viewer scenarios.

Implementation#

The benchmark is implemented in benchmarks/brax/brax_envs/time_envs_device.ipynb. By running the notebook, the following files are generated.

  • report_{env_name}_{gpu_name}.json: Hardware specifications and timing results

  • progress_{env_name}_{gpu_name}.csv: Detailed training progress data

  • consolidated_benchmark_comparison.png: Visual comparison of all runs

Contents