[Real-world active learning] PR 4: in-process RealRobot + babyrobot submodule

[amburger66]

Plan PR 4 of claude_plans/BRIDGE_INTERFACE_PLAN.md (in BabyRobotPredicator).

Cross-link: this is the predicators half of BabyRobotPredicator#49 — PR 2: RealRobot in-process interface, which added the RealRobot library this PR calls. Per CONTRIBUTING.md, the two are one change split across repos.

What

Replaces the two-process JSON socket bridge with an in-process library call. No behavior change on the arm — still whole-episode, still open-loop.

• Submodule. BabyRobotPredicator is added as submodules/BabyRobotPredicator, pinned at 3565015 (its plan PRs 0–2: pyproject.toml, the message dataclasses, and RealRobot).
• real_robot_bridge.py. RealRobotBridgeClient and the entire socket/JSON transport are deleted — no sockets, no serialization, no length-prefix framing, no remote-code-execution surface, because there is no longer a wire. What remains is:
  • make_real_robot(), a factory that imports babyrobot inside the function body and constructs the RealRobot with dry / perception from config;
  • _split_actions(actions, layout), now driven by a GripperJointLayout (four numbers) rather than reaching into a SingleArmPyBulletRobot;
  • split_actions_by_option(actions), which groups a buffered episode into per-option chunks. Unused by this PR — it is what the real-world wrapper (plan PR 6) needs to ask for one observation per option boundary — but it lands and is tested here alongside its sibling.
• pybullet_domino_real.py. Holds a RealRobot instead of a socket client and calls it directly. Same buffering, same whole-episode flush, same homing before the first option.
• Settings. real_robot_bridge_host / _port are dropped (there is no socket to address); real_robot_dry and real_robot_perception are added.

babyrobot stays optional

It is a private repo, so predicators' CI cannot clone it. That constraint is enforced, not assumed:

• not in install_requires; installed separately from the submodule path;
• no module under predicators/ imports babyrobot at module level — a test asserts this on the parse tree, so it holds whether or not babyrobot happens to be installed;
• mypy.ini gets a [mypy-babyrobot.*] stanza covering both the present and absent cases;
• asking for real-robot execution without it raises one clear error naming the submodule and the two install commands, and only when the factory is actually called;
• README.md documents the optional install.

Tooling scoped away from submodules/

run_autoformat.sh reformatted 40+ files inside the submodule on its first run — dirtying another repo's working tree with this repo's style config. run_autoformat.sh, scripts/run_checks.sh (which the README points at as the one-line local check), conftest.py, .coveragerc and mypy.ini are all scoped to skip submodules/ now.

The submodule lives at submodules/, not third_party/, to keep it distinct from the pre-existing predicators/third_party/ — that holds vendored source (fast_downward_translator, ikfast) imported through the predicators namespace, whereas this is a pointer to a live external repo installed as its own distribution and imported as babyrobot.

Not in this PR

Per the plan these are PRs 5–7 and are deliberately absent: observation conversion (state_from_observation / task_from_observation), option-boundary observation, twin resynchronization, and RealWorldEnv.

How this was tested

• tests/pybullet_helpers — 64 passed.
• tests/pybullet_helpers/test_real_robot_bridge.py collects 15 tests in both configurations (so a silent mass-skip cannot pass for green): 15 pass with the submodule installed, 7 pass / 8 skip without it. import predicators.envs works in both.
• mypy . --config-file mypy.ini — clean, 661 source files.
• Full pytest . --pylint -m pylint gate — the changed files rate 10.00/10; yapf --diff clean.
• Dry-robot smoke check (no hardware): a Pick chunk and a Place chunk shipped as separate step calls produce exactly one close at the hand. This is the bug that forced whole-episode shipping — the split still re-emits the leading close on the Place, and RealRobot's session-wide gripper dedup drops it, so the release is not dropped.

Not validated on hardware. The Franka was not moved; every check above ran with RealRobot(dry=True) or against a stub. The arm smoke test is human-run at the rig (plan verification step 9).

Note for reviewers

.gitmodules uses the SSH URL (git@github.com:BasisResearch/BabyRobotPredicator.git), matching how the repo is cloned today. Say the word if an HTTPS URL suits more people.

🤖 Generated with Claude Code

[yichao-liang]

Thanks for the great work!
One comment: there is already a third_party directory inside the predicators directory here. Would it make sense to install the submodule there?

[amburger66]

Thanks for the great work! One comment: there is already a third_party directory inside the predicators directory here. Would it make sense to install the submodule there?

@yichao-liang Thanks for reviewing! Right, I had the same thought so I asked claude --

predicators/third_party/ is vendored source — 42 files checked into this repo as ordinary files (fast_downward_translator, ikfast),
imported through the predicators namespace: from predicators.third_party.fast_downward_translator.translate import ...
(predicators/utils.py:71).

third_party/BabyRobotPredicator is a pointer to a live external repo, installed as its own pip distribution from its own
pyproject.toml and imported as babyrobot — never as predicators.third_party.babyrobot. Its location on disk doesn't determine its
import path; pip install -e third_party/BabyRobotPredicator does.

Reasons that distinction pushed it to the top level:

1. predicators/ is the packaged directory. setup.py runs find_packages(include=["predicators", "predicators.*"]) with
include_package_data=True. Dropping a second repo's entire tree — its .git, scenes, models, its own tests — inside the directory that
pip install -e . treats as the package root invites it into predicators' own distribution. The existing vendored dirs are small,
deliberately-chosen source files, not a whole checkout.
2. Nesting would imply a namespace relationship that doesn't exist. Everything currently under predicators/third_party/ is reachable
as predicators.third_party.*. babyrobot isn't, and shouldn't be — the shared message classes are the contract precisely because both
repos import the same babyrobot.realrobot.messages.

I think this makes sense, but maybe we want to name the folder something else (like submodules) to disambiguate. Thoughts?

[yichao-liang]

I see! submodules sounds like a better name, let’s use that!