Performance Benchmarks¶

KELab benchmarks run headless on an NVIDIA GeForce RTX 4090 and Intel Core i7-13700K with 4096 environments. FPS counts environment control steps.

KangEngine revision: cef3dd2.

Simulation and Training Throughput¶

Task	Framework	Environments	Environment Step FPS	Step + Inference FPS	Step + Inference + Train FPS
Cartpole	KELab	4096	1,153,317	914,046	527,070
Cartpole	IsaacLab v2.3.2 (local)	4096	978,112	794,296	491,239
G1 Rough	KELab	4096	104,009	95,889	84,954
G1 Rough	IsaacLab v2.3.2 (local)	4096	85,039	85,122	76,552

Cartpole: RL-Games, 50 training iterations.
G1 Rough: RSL-RL 5.4.2, 50 training iterations; environment-only FPS is measured separately over 100 steps with random actions.
Following IsaacLab v2.3.2 benchmark scripts, values are arithmetic means of native per-iteration FPS (per-step FPS for non-RL), with no warm-up exclusion or added CUDA synchronization. Inference includes rollout collection.
Local IsaacLab runs use the fresh pinned reference environment on the same GPU/CPU, with matching seed 42 and run lengths. These are single runs; Torch/Python versions and simulation implementations differ between frameworks.

For reference, IsaacLab v2.3.2 publishes the following RTX 4090 results with a Ryzen 9 7950X CPU. CPU and measurement settings differ from the local runs. Its published table does not specify the training iteration count; the local measurement length is given above.

IsaacLab Environment	Environments	Environment Step FPS	Step + Inference FPS	Step + Inference + Train FPS
Isaac-Cartpole-Direct-v0	4096	1,100,000	910,000	510,000
Isaac-Velocity-Rough-G1-v0	4096	94,000	88,000	82,000

Humanoid Training Time¶

RL-Games, 4096 environments × 32 rollout steps × 500 iterations: 65,536,000 environment steps on RTX 4090.

RL Library	KELab — Torch (s)	KELab — Warp base (s)	IsaacLab v2.3.2 Published Time (s)
RL-Games	193.80	176.65	201

Both KELab compute backends were measured on 2026-09-09 with Torch compilation disabled. Times are script-reported training durations. The IsaacLab reference matches GPU model and step count; other system and task settings may differ.

G1 Rough: 4000-Iteration Training¶

Both frameworks ran locally with RSL-RL 5.4.2, matching PPO settings, seed 42, and 4096 environments × 24 rollout steps × 4000 iterations: 393,216,000 environment steps. KELab uses the reference USD robot with 37 DOFs, three convex hulls and implicit PD.

Metric	KELab	IsaacLab v2.3.2
Rollout collection time (s)	3,514.75	3,927.76
PPO update time (s)	452.50	445.07
Total measured training time (s)	3,967.25	4,372.83
Full-run training FPS	99,116	89,923
Final mean episode reward	29.72	26.75
Final mean episode length (s)	19.69	19.71

Training FPS is total steps divided by summed collection and update times, excluding initialization and checkpoint saving. Reward and episode length are averages over the final 50 iterations. Results are from one seed; Python, Torch and physics backends differ between frameworks.

Benchmark Commands¶

The KELab repository also provides scripts/benchmarks/isaaclab_reference/ with pinned dependencies and standalone IsaacLab v2.3.2 train/play commands. Its README describes a fresh conda setup using Isaac Sim 5.1.0 and RSL-RL 5.4.2.

Run from the KELab repository:

# Cartpole throughput during RL-Games training
python scripts/benchmarks/benchmark_rl_games.py \
  --task Cartpole-Direct-v0 --num_envs 4096 --max_iterations 50 \
  --output logs/benchmarks/cartpole.json

# G1 rollout and training throughput
python scripts/benchmarks/benchmark_rsl_rl.py \
  --task G1-Velocity-Rough-v0 --num_envs 4096 --max_iterations 50 \
  --output logs/benchmarks/g1_training.json

# G1 environment-only throughput
python scripts/benchmarks/benchmark_env.py \
  --task G1-Velocity-Rough-v0 --num_envs 4096 --num_frames 100 \
  --output logs/benchmarks/g1_env.json

# Humanoid training time
python scripts/rl_games/train.py --task Humanoid-v0 \
  --num_envs 4096 --device cuda:0 --compute_backend torch \
  --max_iterations 500

# Humanoid training time — Warp base
python scripts/rl_games/train.py --task Humanoid-v0 \
  --num_envs 4096 --device cuda:0 --compute_backend warp \
  --max_iterations 500

# G1 Rough training
python scripts/rsl_rl/train.py --task G1-Velocity-Rough-v0 \
  --num_envs 4096 --device cuda:0 --seed 42 --max_iterations 4000