Skip to content

[DEMO] Golden policy correctness check - #12

Open
angehu-nv wants to merge 4 commits into
perf-smoke/develop-stagingfrom
angehu-test/golden-ci-dryrun
Open

[DEMO] Golden policy correctness check#12
angehu-nv wants to merge 4 commits into
perf-smoke/develop-stagingfrom
angehu-test/golden-ci-dryrun

Conversation

@angehu-nv

Copy link
Copy Markdown

Demo only; do not merge.

Reject an unknown or misspelled physics/render backend in tasks.json at
load time instead of silently producing a backend_key that no threshold
targets (which disables the FPS gate) or emitting a bad Hydra presets=
token that crashes the GPU benchmark job at launch.
Runs a frozen policy checkpoint per (task, backend) after the perf smoke
benchmark, on the same warm runner via the shared isaac-rollout composite
action, and scores behavioural KPIs (reward / episode length / success rate)
against hard-set thresholds in golden_tasks.json. Standalone from perf (no
baselines); advisory via a separate gate_config.golden_blocking toggle. Golden
piggybacks perf cells so it needs no extra image pull; a separate ubuntu-latest
golden_aggregate emits its own omni-github artifact (test_tool_id=golden-policy).
Throwaway branch (do not merge). Trims tasks.json to Cartpole/physx (golden
pair) + Velocity-Flat-G1/physx (no golden -> must skip); golden_tasks.json to
Cartpole/physx. Adds a TEST-only step that pulls a real rsl_rl checkpoint from
the fork release 'angehu-test-golden-ckpt' into a workspace GOLDEN_ROOT and
times it, so the real --checkpoint pull+load path is exercised on RTX 6000.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant