Skip to content

Remove the phased agent code - #190

Merged
yichao-liang merged 1 commit into
masterfrom
empiric-phased-cleanup
Sep 28, 2026
Merged

yichao-liang merged 1 commit into
masterfrom
empiric-phased-cleanup

Conversation

@yichao-liang

Copy link
Copy Markdown
Collaborator

Summary

Stacked on #189 (base empiric-configs); review and merge after it.

The EMPIRIC agent and its six baselines all run under the continual protocol, where each arm plays the levels of one environment through ContinualPlayMixin.play_level.
This deletes the phased experiment path (learn sessions between online learning cycles, solve sessions per test task) and the code only that path reached: 70 files, about 12.8k lines.

Deleted:

  • AgentModelBasedApproach, the agent explorers and their base class, and the subgoal_annotations execution monitor;
  • SamplerLearningMixin and the sampler synthesis tools;
  • the NL world model and GNN dynamics baselines;
  • the Docker sandbox backend and its agent runner;
  • the learn-notes prompt templates and their goldens;
  • in AgentSimLearningApproach: the methods only the phased path called, the oracle-simulator branches, and the per-cycle fit diagnostics with their fields;
  • the settings flags that lost their last reader.

Kept:

  • AgentModelFreeApproach keeps the option model, tool-context sync and checkpointing that the continual arms inherit. Its _solve now raises, since no arm solves tasks outside play_level.
  • ExoPredicator-era code, the arm classes outside the benchmark, the env variants and the real-robot stack. No deleted file exists at the ExoPredicator tag or at the upstream fork point.

Sketch refinement now draws a step's params from its ground sampler, or uniformly; the learned parameterized samplers were phased-only.

empiric/common.yaml and envs.yaml stop setting the phased pipeline's flags (make_*_videos, num_online_learning_cycles, max_num_steps_interaction_request, the online early-stopping flags, online_nsrt_learning_requests_per_cycle).
Under the continual protocol run_pipeline returns right after run_continual, so none of them was read.

Tests of deleted code are deleted.
test_execution_monitoring.py returns to its upstream version, and the interaction-request bookkeeping test now exercises the online process learner, the remaining implementer of that hook.

Test plan

  • yapf, isort 5.10.1, docformatter, mypy and pylint on this commit (sbatch replay of CI's static jobs).
  • The 8 CI shards in the CI container.

🤖 Generated with Claude Code

The EMPIRIC agent and its baselines all run under the continual protocol.
This removes the phased experiment path (learn sessions between online
learning cycles, solve sessions per test task) and the code only it
reached:

- AgentModelBasedApproach, the agent explorers, the subgoal-annotation
  execution monitor and the sampler learning mixin
- the NL world model and GNN dynamics baselines
- the Docker sandbox backend and the sampler synthesis tools
- the learn-note templates, the oracle-simulator branches, the per-cycle
  fit diagnostics, and the settings flags nothing reads anymore

AgentModelFreeApproach keeps the option model, tool-context sync and
checkpointing that the continual arms inherit. Its task solver raises,
since the arms play levels through ContinualPlayMixin.play_level.
ExoPredicator-era code is untouched.
@yichao-liang
yichao-liang changed the base branch from empiric-configs to master September 28, 2026 00:08
@yichao-liang
yichao-liang merged commit 1402d79 into master Sep 28, 2026
14 checks passed
@yichao-liang
yichao-liang deleted the empiric-phased-cleanup branch September 28, 2026 09:32
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant