[BugFix] RSSMRollout not advancing state/belief across time steps in Dreamer

[cmdout]

Description

This PR fixes the RSSM rollout state propagation in RSSMRollout.forward.
At each timestep, after computing the prior and posterior, we now:

• shift ("next", *) keys to current using step_mdp(..., keep_other=True), and
• merge the next step inputs before continuing the rollout.

Concretely:

• add step_mdp import, and
• call step_mdp inside the time loop before selecting in_keys and updating with the next step’s values.

This ensures ("next", "state") and ("next", "belief") are correctly advanced to state and belief for the subsequent step.

Motivation and Context

Previously, the rollout did not propagate state and belief from ("next", ...) back to the current keys, causing the loop to reuse stale state/belief. This broke the intended posterior-to-prior chaining in Dreamer’s RSSM and led to incorrect latent rollouts (and potentially training instability and degraded reconstruction/reward predictions).

The fix aligns the implementation with Dreamer’s design and the world model wiring used elsewhere (e.g., in dreamer_utils._dreamer_make_world_model), where the prior for step t+1 must consume the posterior outputs of step t.

If there is an open issue tracking this bug, please link it here.
Example: close #XXXX

• I have raised an issue to propose this change (required for new features and bug fixes)

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds core functionality)
• Breaking change (fix or feature that would cause existing functionality to change)
• Documentation (update in the documentation)
• Example (update in the folder of examples)

Checklist

• I have read the CONTRIBUTION guide (required)
• My change requires a change to the documentation.
• I have updated the tests accordingly (required for a bug fix or a new feature).
• I have updated the documentation accordingly.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/rl/3236

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 14 New Failures, 1 Cancelled Job, 5 Unrelated Failures

As of commit 601698f with merge base 8570c25 ():

NEW FAILURES - The following jobs have failed:

• Continuous Benchmark (PR) / CPU Pytest benchmark (gh)
  Workflow failed! Resource not accessible by integration
• Habitat Tests on Linux / tests (3.9, 12.8) / linux-job (gh)
  RuntimeError: Command docker exec -t b47a9debc0f4fc12b84e265c8be0b4043f5a850dd222291bfddcc9a23464ac9f /exec failed with exit code 1
• Libs Tests on Linux / unittests-gym (3.9, 12.8) / linux-job (gh)
  RuntimeError: SumSegmentTreeFp32 is not available. See warning above.
• Push Binary Nightly / build-wheel-unix (linux, ubuntu-22.04, 3.9, cp39-cp39, cpu, cpu) (gh)
• SOTA Tests on Linux / tests (3.9, 12.8) / linux-job (gh)
  RuntimeError: Command docker exec -t 161875347bcb993c09c40c10d56d4b8e5b349fff51956847253cd677103372c6 /exec failed with exit code 1
• Unit-tests on Linux / tests-cpu (3.10) / linux-job (gh)
  test/test_modules.py::TestDreamerComponents::test_rssm_rollout[6-4-30-20-batch_size2-device0]
• Unit-tests on Linux / tests-cpu (3.11) / linux-job (gh)
  test/test_modules.py::TestDreamerComponents::test_rssm_rollout[6-4-30-20-batch_size2-device0]
• Unit-tests on Linux / tests-cpu (3.12) / linux-job (gh)
  test/test_modules.py::TestDreamerComponents::test_rssm_rollout[6-4-30-20-batch_size2-device0]
• Unit-tests on Linux / tests-cpu (3.9) / linux-job (gh)
  test/test_modules.py::TestDreamerComponents::test_rssm_rollout[6-4-30-20-batch_size2-device0]
• Unit-tests on Linux / tests-gpu (3.11, 12.8) / linux-job (gh)
  test/test_modules.py::TestDreamerComponents::test_rssm_rollout[6-4-30-20-batch_size2-device0]
• Unit-tests on Linux / tests-olddeps (3.9, 11.6) / linux-job (gh)
  test/test_modules.py::TestDreamerComponents::test_rssm_rollout[6-4-30-20-batch_size2-device0]
• Unit-tests on Linux / tests-optdeps (3.11, 12.8) / linux-job (gh)
  test/test_modules.py::TestDreamerComponents::test_rssm_rollout[6-4-30-20-batch_size2-device0]
• Unit-tests on Linux / tests-stable-gpu (3.10, 11.8) / linux-job (gh)
  test/test_modules.py::TestDreamerComponents::test_rssm_rollout[6-4-30-20-batch_size2-device0]
• Unit-tests on Windows / unittests-cpu (3.10, windows.4xlarge, cpu) / windows-job (gh)
  test/test_modules.py::TestDreamerComponents::test_rssm_rollout[6-4-30-20-batch_size2-device0]

CANCELLED JOB - The following job was cancelled. Please retry:

• Continuous Benchmark (PR) / GPU Pytest benchmark (gh)
  ##[error]The operation was canceled.

BROKEN TRUNK - The following jobs failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• Push Binary Nightly / build-wheel-unix (macos, macos-latest, 3.9, cp39-cp39, cpu, cpu) (gh) (trunk failure)
  Process completed with exit code 1.
• Push Binary Nightly / build-wheel-windows (3.9, 3.9) (gh) (trunk failure)
  Process completed with exit code 1.
• Push Binary Nightly / test-wheel-unix (linux, ubuntu-22.04, 3.9, cp39-cp39, cpu, cpu) (gh) (trunk failure)
  Process completed with exit code 1.
• Push Binary Nightly / test-wheel-unix (macos, macos-latest, 3.9, cp39-cp39, cpu, cpu) (gh) (trunk failure)
  Process completed with exit code 1.
• Push Binary Nightly / test-wheel-windows (3.9, 3.9) (gh) (trunk failure)
  Process completed with exit code 1.

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[meta-cla]

Hi @cmdout!

Thank you for your pull request and welcome to our community.

Action Required

In order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at [email protected]. Thanks!

[vmoens]

Hello @cmdout
thanks for this! Looks ok but do you have a training curve to show? Dreamer has been broken for quite some time now and I did not find the time to refresh it, if this fix helps us training it better it'd be awesome!
Also can you sign the CLA?