Include run_name in training logs for better run traceability

[ShaoshuSu]

Description

Previously, during training, the on_policy_runner logging output did not display the current run name, making it harder to distinguish logs between different training sessions when multiple runs were launched in succession or in parallel.

To improve the training log clarity and traceability, I added support to display the configured run_name in the logging output of rsl_rl/runners/on_policy_runner.py.

If the run_name field is specified in the configuration—such as in the Isaac Lab training example:

./isaaclab.sh -p scripts/reinforcement_learning/rsl_rl/train.py \
    --task Isaac-Cartpole-v0 --run_name cartpole_rsl_rl

—the run name will now appear in each log update under the label "Run name:":

################################################################################
                        Learning iteration 5/150                        

                          Run name: cartpole_rsl_rl
                       Computation: 90014 steps/s (collection: 0.668s, learning 0.060s)
             Mean action noise std: 0.96
          Mean value_function loss: 0.2114
               Mean surrogate loss: -0.0055
                 Mean entropy loss: 1.3755
                       Mean reward: -3.72
               Mean episode length: 91.66
              Episode_Reward/alive: 0.2862
        Episode_Reward/terminating: -0.0051
           Episode_Reward/pole_pos: -0.9367
           Episode_Reward/cart_vel: -0.0072
           Episode_Reward/pole_vel: -0.0056
      Episode_Termination/time_out: 11.0000
Episode_Termination/cart_out_of_bounds: 36.1875
--------------------------------------------------------------------------------
                   Total timesteps: 393216
                    Iteration time: 0.73s
                      Time elapsed: 00:00:04
                               ETA: 00:01:55

This enhancement helps users quickly identify the corresponding training session, which is especially useful when monitoring multiple or long-running experiments.

This addition is optional and non-intrusive: if run_name is not set, the log display remains unchanged.

What’s Changed

• Display of run_name added to training logs in on_policy_runner.py.
• run_name is conditionally fetched from self.cfg and formatted neatly using the same padded alignment style as other log fields.
• Logging behavior remains unchanged if run_name is not defined.

Type of Change

• 🧠 Enhancement (non-breaking addition of functionality)

Checklist

• Ran pre-commit run --all-files and all pre-commit checks pass
• Verified that the training output now includes the run_name when provided
• Confirmed no unintended side effects or regressions in log formatting
• Maintained backward compatibility (log remains unchanged if run_name is not set)
• No new warnings introduced

[ClemensSchwarke]

What do you think @Mayankm96 ?

[epalmaEth]

@ClemensSchwarke we need a Logger class asap 🔥

[epalmaEth]

What about using --task for the run name instead of an additional --run_name parameter? Also i think we need to update the config example with the additional name.

[ClemensSchwarke]

run_name actually already exists. It is just the time stamp by default but can be passed as a CLI argument.

[epalmaEth]

Does it make sense to use run_name and not task still?

[ClemensSchwarke]

I think so because task would usually be the same for all your trainings. If you want to differentiate them you would need run_name.