Release the SpeedTuning simulation reproduction

[DaivdYuan]

Summary

This PR publishes the official simulation reproduction for SpeedTuning: Speeding Up Policy Execution with Lightweight Reinforcement Learning.

• Add complete MuJoCo tasks for pick-and-place, insertion, and tea-bag manipulation.
• Add bundled scripted base policies so every supported experiment can run from scratch without pretrained artifacts.
• Add task-specific Rainbow DQN speed-policy training, periodic snapshots, evaluation, and fixed-speed frontier sweeps.
• Implement decision-level temporal acceleration with fresh receding-horizon action chunks and shared training/evaluation timing semantics.
• Add external action-chunk policy adapters, retained ACT checkpoint loading, and configurable state, visual, or external speed observations.
• Add archival paper and ablation manifests plus fully runnable scripted reproduction presets.
• Add seeded randomized tea-bag evaluation while preserving the fixed-pose historical environment.
• Add machine-readable reference results, a detailed reproduction guide, and external-policy integration documentation.
• Add clean Python packaging, console entry points, pinned physics dependencies, lockfile, citation metadata, licenses, attribution, contribution guidance, and CI.
• Replace the placeholder README with the complete ICRA 2025 research release, including the project page, arXiv:2608.09138, conference paper, video, teaser, citation, and hardware-neutral setup.

Public reproduction surface

The release supports the complete simulator loop:

1. execute each task with the bundled scripted policy;
2. train a task-specific speed controller from scratch;
3. evaluate learned and fixed-speed policies on matched seeds;
4. report physical acceleration from executed MuJoCo steps;
5. integrate external action-chunk task policies through a stable adapter.

No trained checkpoints are committed or required.

Reference results

One seeded 100,000-decision run per task produced:

Protocol	Adaptive SpeedTuning	Matched fixed speed
Pick-and-place	98% at 3.856x	66% at 3.846x
Insertion	97% at 2.387x	52% at 2.381x
Tea bag, randomized poses	78% at 2.077x	24% at 2.075x

The versioned JSON record includes presets, training seeds, held-out seed ranges, episode counts, and metric definitions.

Validation

• uv lock --check passes with Python 3.10.
• Full optional dependency installation succeeds from the clean public branch.
• MUJOCO_GL=egl uv run pytest -q: 50 passed.
• All three scripted tasks succeed at nominal speed and at 1.5x.
• Chunked-policy integration succeeds for all three tasks.
• Rainbow proof-of-concept optimization passes with finite losses and parameter updates.
• The documented 1,000-decision CPU smoke train completes with 894 optimizer updates and a loadable checkpoint.
• Evaluation and fixed-speed sweep commands complete successfully.
• Source and wheel distributions build successfully and pass twine check.
• The wheel installs in an isolated environment and runs the packaged tea-bag simulator outside the checkout.
• Packaged metadata contains the public repository and arXiv links.
• CITATION.cff validates against CFF schema 1.2.0.
• GitHub-flavored README rendering resolves the title, badge, teaser, arXiv link, and results section.
• Project page, arXiv abstract/PDF, conference PDF, video, and DOI links resolve successfully.
• The staged tree contains no model artifacts, machine-local paths, or private repository references.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.