Update the template generator for RSL-RL

[ClemensSchwarke]

Description

This PR adds Distillation to the list of available algorithms for RSL-RL. It also updates the config templates to RSL-RL version 5.0.

Type of change

• Template generator update

Checklist

• I have run the pre-commit checks with ./isaaclab.sh --format
• My changes generate no new warnings

[greptile-apps]

Greptile Summary

This PR updates the Isaac Lab template generator to target RSL-RL v5.0 by introducing a new DISTILLATION algorithm template and migrating the existing PPO template from the old RslRlPpoActorCriticCfg API to the new RslRlMLPModelCfg / obs_groups model. It also improves the CLI display of algorithms in the RL library table (word-wrapping with textwrap.fill and overflow="fold") and updates the RSL-RL distributed training flag to Yes.\n\nKey changes:\n- tools/template/common.py — "DISTILLATION" added to SINGLE_AGENT_ALGORITHMS; the generator auto-discovers it via the new template file.\n- tools/template/templates/agents/rsl_rl_ppo_cfg — Migrated to RSL-RL v5.0 API: RslRlPpoActorCriticCfg → RslRlMLPModelCfg, with explicit obs_groups and separate actor/critic model definitions.\n- tools/template/templates/agents/rsl_rl_distillation_cfg (new) — Provides a starter config for the RSL-RL distillation workflow. The teacher model carries init_std=0.0 in its GaussianDistributionCfg, which is a potentially degenerate initialisation (zero standard deviation) that should be verified against RSL-RL's implementation.\n- tools/template/cli.py — Table display improvements and RSL-RL distributed-training flag updated to State.Yes.

Confidence Score: 4/5

Safe to merge after confirming how RSL-RL v5.0 handles init_std=0.0 on the teacher's Gaussian distribution.

All changes are confined to template generator tooling (not runtime Isaac Lab code) and are straightforward. The one substantive concern is the init_std=0.0 value on the teacher model in rsl_rl_distillation_cfg. If RSL-RL v5.0 treats this as deterministic/no noise, the template is correct; if it feeds this directly into log(std) or similar, the generated config will produce NaN at runtime for any user who runs distillation without modifying the template.

tools/template/templates/agents/rsl_rl_distillation_cfg — specifically the init_std=0.0 value on the teacher model config.

Important Files Changed

Filename	Overview
tools/template/cli.py	Adds textwrap.fill import for word-wrapping algorithm names in the table, switches column overflow to "fold", and updates RSL-RL's distributed training support to State.Yes for v5.0.
tools/template/common.py	Adds "DISTILLATION" to SINGLE_AGENT_ALGORITHMS so it is discoverable and validated by the generator's template-scanning logic.
tools/template/templates/agents/rsl_rl_distillation_cfg	New RSL-RL Distillation config template; init_std=0.0 on the teacher's Gaussian distribution is potentially degenerate (zero std). Also carries a stale 2022-2025 copyright year despite being created in 2026.
tools/template/templates/agents/rsl_rl_ppo_cfg	Updated PPO template for RSL-RL v5.0: replaces RslRlPpoActorCriticCfg with the new RslRlMLPModelCfg API, adds obs_groups, and separates actor/critic into distinct model configs.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[User runs CLI tool] --> B{Single-agent workflow?}
    B -- Yes --> C[Discover RL libraries\nand algorithms via\nget_algorithms_per_rl_library]
    C --> D[Display updated table\nwith textwrap.fill + overflow=fold]
    D --> E[User selects: rsl_rl]
    E --> F[Available algorithms\nPPO / DISTILLATION]
    F --> G{User picks algorithm}
    G -- PPO --> H[Render rsl_rl_ppo_cfg\nusing new RslRlMLPModelCfg API\nobs_groups + actor + critic]
    G -- DISTILLATION --> I[Render rsl_rl_distillation_cfg\nstudent init_std=1.0\nteacher init_std=0.0 ⚠️]
    H --> J[Write .py config to agents/]
    I --> J

Loading

_{Reviews (1): Last reviewed commit: "update template generator for rsl rl" | Re-trigger Greptile}

[isaaclab-review-bot]

🤖 Isaac Lab Review Bot

Summary

This PR updates the RSL-RL template generator to support the v5.0 API: migrates the PPO template from the deprecated RslRlPpoActorCriticCfg (policy field) to the new RslRlMLPModelCfg pattern (separate actor/critic with obs_groups), adds a new Distillation algorithm template, updates the CLI table formatting to handle wider algorithm lists, and marks RSL-RL as supporting distributed training. The changes are correct and well-aligned with the existing rl_cfg.py and distillation_cfg.py config classes.

Architecture Impact

No cross-module impact — changes are self-contained within tools/template/. These template files are only consumed by the template generator CLI (tools/template/cli.py → generator.py) to scaffold new projects. Existing projects and runtime code are unaffected.

The SINGLE_AGENT_ALGORITHMS list in common.py is used by get_algorithms_per_rl_library() in generator.py to validate discovered algorithms. Adding DISTILLATION there is necessary and correct — the function will only associate it with rsl_rl since only rsl_rl_distillation_cfg exists as a template file.

Implementation Verdict

Minor fixes needed — Correct approach, 1 issue to address (license header year).

Test Coverage

No tests required. The template generator has no existing test suite, and these are template/config/tooling changes that affect only scaffolded project output. The risk of regression is low since template correctness is validated by pre-commit checks on generated output.

CI Status

• ✅ pre-commit — passed
• ✅ Check secrets — passed
• ✅ labeler — passed
• ⏳ Build Docs, license-check, test-general, test-isaaclab-tasks — in progress

Findings

🔵 Improvement: templates/agents/rsl_rl_distillation_cfg:1 and templates/agents/rsl_rl_ppo_cfg:1 — License header year is 2022-2025 but the rest of the codebase has been updated to 2022-2026.

The new distillation template copies the stale year from the existing PPO template. Since this PR touches the PPO template anyway, both should be updated to 2022-2026 for consistency. (The pre-commit hook passed, so this isn't enforced by CI, but it's still the correct year per the rest of the repo.)

[ooctipus]

is this target to main or develop?