[KraHsu]

Problem

genelab export <rsl_rl-task> <ckpt> -f torchscript|onnx failed for every
rsl_rl checkpoint (e.g. GeneLab-Inverted-Pendulum-v0) with:

backend 'rsl_rl' did not expose an actor module; export is not supported for this checkpoint type

(This is the next failure after PR #156 fixed the separate "No display detected" crash.)

Root cause

_extract_rsl_rl_actor (src/genelab/rl/backends/rsl_rl.py) read
runner.alg.actor_critic and returned (None, num_obs) when missing. The
installed rsl_rl's PPO.__init__ no longer has actor_critic — it sets
self.actor, self._raw_actor, self.critic, self._raw_critic directly
(verified via inspect.getsource). So the attribute path was stale,
actor_module ended up None, and _export.py raised.

Fix

Source the actor from alg._raw_actor (the uncompiled module; falls back to
alg.actor, which may be a torch.compile wrapper). The actor is an
MLPModel whose own forward takes a TensorDict and self-normalizes, so we
use its as_jit() wrapper — _TorchMLPModel, which exposes the flat
forward(obs) -> deterministic action contract (learned obs normalizer baked
in) that GeneLab's exporter wraps with env-side scale/clip. Input dim comes
from actor.obs_dim. The legacy alg.actor_critic.actor / act_inference
path is kept as a fallback for older rsl_rl.

Verification (headless, real checkpoint)

QD_GRAPH=0 QT_QPA_PLATFORM=offscreen genelab export GeneLab-Inverted-Pendulum-v0 <ckpt> -f {torchscript,onnx} --out ...

• both exit 0 and write .ts / .onnx + .metadata.json
• TorchScript loads and runs: obs_dim 8 -> action_dim 1
• ONNX passes onnx.checker; IO obs:[batch,8] -> actions:[batch,1] (dynamic batch)

Tests

tests/test_export.py:

• test_extract_rsl_rl_actor_new_layout — builds a real rsl_rl MLPModel,
  asserts extraction returns a flat-input deterministic module with the right
  dim, and runs it through the full export_policy -> loadable .ts.
• test_extract_rsl_rl_actor_legacy_layout — alg.actor_critic.actor path.
• test_extract_rsl_rl_actor_missing_returns_none — (None, num_obs) so the
  caller raises a clear error.

tests/test_export.py — 10 passed; export+rsl suite 17 passed; ruff clean.

Docs

docs/concepts/eval-and-export.{en,zh}.md updated to describe the new
extraction path (convention: docs update for rl/backends/ changes).

Reference audit

find_referencing_symbols(_extract_rsl_rl_actor) → only caller is
RslRlBackend.make_inference_setup (rsl_rl.py:120); signature unchanged, no
other caller affected.

🤖 Generated with Claude Code