Fixes benchmark scripts when tensorboard logs are missing

scripts/benchmarks/benchmark_rlgames.py

@@ -103,13 +103,9 @@
 from scripts.benchmarks.utils import (
     get_backend_type,
     get_preset_string,
-    get_success_rate_log,
     log_app_start_time,
-    log_convergence,
     log_python_imports_time,
-    log_rl_policy_episode_lengths,
-    log_rl_policy_rewards,
-    log_rl_policy_success_rates,
+    log_rl_training_metrics,
     log_runtime_step_times,
     log_scene_creation_time,
     log_simulation_start_time,
@@ -288,15 +284,12 @@ def main(
         log_simulation_start_time(benchmark, Timer.get_timer_info("simulation_start") * 1000)
         log_total_start_time(benchmark, (task_startup_time_end - app_start_time_begin) / 1e6)
         log_runtime_step_times(benchmark, rl_training_times, compute_stats=True)
-        log_rl_policy_rewards(benchmark, log_data["rewards/iter"])
-        log_rl_policy_episode_lengths(benchmark, log_data["episode_lengths/iter"])
-        success_rates = get_success_rate_log(log_data)
-        if success_rates is not None:
-            log_rl_policy_success_rates(benchmark, success_rates)
-        log_convergence(
+        log_rl_training_metrics(
             benchmark,
-            log_data["rewards/iter"],
-            args_cli.task,
+            log_data,
+            reward_tag="rewards/iter",
+            episode_length_tag="episode_lengths/iter",
+            task=args_cli.task,
             workflow="rl_games",
             should_check_convergence=args_cli.check_convergence,
             reward_threshold=args_cli.reward_threshold,

scripts/benchmarks/benchmark_rsl_rl.py

@@ -105,13 +105,9 @@
 from scripts.benchmarks.utils import (
     get_backend_type,
     get_preset_string,
-    get_success_rate_log,
     log_app_start_time,
-    log_convergence,
     log_python_imports_time,
-    log_rl_policy_episode_lengths,
-    log_rl_policy_rewards,
-    log_rl_policy_success_rates,
+    log_rl_training_metrics,
     log_runtime_step_times,
     log_scene_creation_time,
     log_simulation_start_time,
@@ -287,16 +283,12 @@ def main(
         log_simulation_start_time(benchmark, Timer.get_timer_info("simulation_start") * 1000)
         log_total_start_time(benchmark, (task_startup_time_end - app_start_time_begin) / 1e6)
         log_runtime_step_times(benchmark, rl_training_times, compute_stats=True)
-        log_rl_policy_rewards(benchmark, log_data["Train/mean_reward"])
-        log_rl_policy_episode_lengths(benchmark, log_data["Train/mean_episode_length"])
-        success_rates = get_success_rate_log(log_data)
-        if success_rates is not None:
-            log_rl_policy_success_rates(benchmark, success_rates)
-
-        log_convergence(
+        log_rl_training_metrics(
             benchmark,
-            log_data["Train/mean_reward"],
-            args_cli.task,
+            log_data,
+            reward_tag="Train/mean_reward",
+            episode_length_tag="Train/mean_episode_length",
+            task=args_cli.task,
             workflow="rsl_rl",
             should_check_convergence=args_cli.check_convergence,
             reward_threshold=args_cli.reward_threshold,

scripts/benchmarks/test/test_training_metrics.py

@@ -0,0 +1,90 @@
+# Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
+# All rights reserved.
+#
+# SPDX-License-Identifier: BSD-3-Clause
+
+"""Unit tests for benchmark training-metric logging helpers."""
+
+from __future__ import annotations
+
+import pytest
+
+from scripts.benchmarks.utils import SUCCESS_RATE_LOG_TAGS, log_rl_training_metrics
+
+
+class _FakeBenchmark:
+    """Collect benchmark measurements without initializing benchmark backends."""
+
+    def __init__(self):
+        self.measurements: list[tuple[str, str, object, str]] = []
+
+    def add_measurement(self, phase, measurement):
+        self.measurements.append((phase, measurement.name, measurement.value, getattr(measurement, "unit", "")))
+
+    def measurement_by_name(self, name: str):
+        return next(m for m in self.measurements if m[1] == name)
+
+
+@pytest.mark.parametrize(
+    "workflow,reward_tag,episode_length_tag",
+    [
+        ("rl_games", "rewards/iter", "episode_lengths/iter"),
+        ("rsl_rl", "Train/mean_reward", "Train/mean_episode_length"),
+    ],
+)
+def test_log_rl_training_metrics_skips_missing_short_run_scalars(
+    workflow: str, reward_tag: str, episode_length_tag: str, capsys: pytest.CaptureFixture[str]
+):
+    """Short benchmark runs may finish before reward and episode-length scalars are emitted."""
+    benchmark = _FakeBenchmark()
+
+    log_rl_training_metrics(
+        benchmark,
+        log_data={},
+        reward_tag=reward_tag,
+        episode_length_tag=episode_length_tag,
+        task="Isaac-Ant-v0",
+        workflow=workflow,
+        should_check_convergence=True,
+    )
+
+    assert benchmark.measurements == []
+    output = capsys.readouterr().out
+    assert f"TensorBoard log is missing '{reward_tag}'" in output
+    assert f"TensorBoard log is missing '{episode_length_tag}'" in output
+    assert f"Cannot check convergence because '{reward_tag}' was not logged" in output
+
+
+@pytest.mark.parametrize(
+    "workflow,reward_tag,episode_length_tag",
+    [
+        ("rl_games", "rewards/iter", "episode_lengths/iter"),
+        ("rsl_rl", "Train/mean_reward", "Train/mean_episode_length"),
+    ],
+)
+def test_log_rl_training_metrics_logs_present_normal_run_scalars(
+    workflow: str, reward_tag: str, episode_length_tag: str, capsys: pytest.CaptureFixture[str]
+):
+    """Normal runs with reward and episode-length scalars should log train metrics."""
+    benchmark = _FakeBenchmark()
+
+    log_rl_training_metrics(
+        benchmark,
+        log_data={
+            reward_tag: [1.0, 2.0, 3.0],
+            episode_length_tag: [10.0, 11.0],
+            SUCCESS_RATE_LOG_TAGS[0]: [0.25, 0.5],
+        },
+        reward_tag=reward_tag,
+        episode_length_tag=episode_length_tag,
+        task="Isaac-Ant-v0",
+        workflow=workflow,
+    )
+
+    assert benchmark.measurement_by_name("Rewards")[2] == [1.0, 2.0, 3.0]
+    assert benchmark.measurement_by_name("Max Rewards")[2] == 3.0
+    assert benchmark.measurement_by_name("Episode Lengths")[2] == [10.0, 11.0]
+    assert benchmark.measurement_by_name("Max Episode Lengths")[2] == 11.0
+    assert benchmark.measurement_by_name("Success Rates")[2] == [0.25, 0.5]
+    assert benchmark.measurement_by_name("success_rate")[2] == 0.5
+    assert "TensorBoard log is missing" not in capsys.readouterr().out

scripts/benchmarks/utils.py

@@ -295,6 +295,52 @@ def log_success(benchmark, tracker, framework_iteration_count: int | None = None
         )
 
 
+def log_rl_training_metrics(
+    benchmark: BaseIsaacLabBenchmark,
+    log_data: dict[str, list[float]],
+    reward_tag: str,
+    episode_length_tag: str,
+    task: str,
+    workflow: str,
+    should_check_convergence: bool = False,
+    reward_threshold: float | None = None,
+    convergence_config: str = "full",
+) -> None:
+    """Log optional RL training metrics from TensorBoard data.
+
+    Short smoke-test runs can finish before the RL framework emits reward or
+    episode-length scalars. Missing tags should skip those measurements instead
+    of failing the whole benchmark.
+    """
+    rewards = log_data.get(reward_tag)
+    episode_lengths = log_data.get(episode_length_tag)
+    if rewards:
+        log_rl_policy_rewards(benchmark, rewards)
+    else:
+        print(f"[WARNING] TensorBoard log is missing '{reward_tag}'; skipping reward benchmark metrics.")
+    if episode_lengths:
+        log_rl_policy_episode_lengths(benchmark, episode_lengths)
+    else:
+        print(f"[WARNING] TensorBoard log is missing '{episode_length_tag}'; skipping episode-length metrics.")
+
+    success_rates = get_success_rate_log(log_data)
+    if success_rates is not None:
+        log_rl_policy_success_rates(benchmark, success_rates)
+
+    if rewards:
+        log_convergence(
+            benchmark,
+            rewards,
+            task,
+            workflow=workflow,
+            should_check_convergence=should_check_convergence,
+            reward_threshold=reward_threshold,
+            convergence_config=convergence_config,
+        )
+    elif should_check_convergence:
+        print(f"[WARNING] Cannot check convergence because '{reward_tag}' was not logged.")
+
+
 def parse_cprofile_stats(
     profile: cProfile.Profile,
     isaaclab_prefixes: list[str],