Reproduction tutorial · August 28, 2026
Instruct-to-Act hierarchical controller
These tutorials test the mechanism behind each paper at a practical scale. They do not reproduce frontier-scale headline results unless explicitly stated.
Source of this tutorial
This tutorial tests the mechanism behind reviewed development 03 in the immutable briefing of August 28, 2026.
Scope and limits
- Reproduction level
- partial mechanism reproduction
- Claim under test
- sparse high-level instructions can guide a faster low-level controller while reducing planner calls.
- Excluded claim
- this tutorial does not reproduce the paper's seven-environment or six-task state-of-the-art comparison.
- Reader, effort, and cost
- advanced reinforcement learning; 6 to 12 hours; CPU or one modest GPU; use a simulator only.
Canonical resources
- Paper: https://arxiv.org/abs/2608.26788 (opens in a new tab)
- Project page: https://zinengtang.github.io/instruct-to-act/ (opens in a new tab)
- Official code: linked on the project page. Record the repository commit and licenses at execution time because they were not copied into this editorial package.
Project plan
| Step | Stage | Effort |
|---|---|---|
| 01 | Choose a small environment | 30 minutes |
| 02 | Controller baseline | 90 minutes |
| 03 | Instruction relabeling | 60 minutes |
| 04 | Language-conditioned controller | 120 minutes |
| 05 | Sparse planner | 60 minutes |
| 06 | Evaluation | 60 minutes |
| 07 | Cleanup | 10 minutes |
01Choose a small environment(30 minutes)
Use a grid-world or simple continuous-control simulator with three goals and deterministic evaluation seeds. Freeze train and test maps.
02Controller baseline(90 minutes)
Train or implement a high-frequency controller that receives observations and acts every step without language. Record success, return, control frequency, and inference latency.
03Instruction relabeling(60 minutes)
Segment successful controller trajectories and label each segment with a templated instruction such as “go to red key” or “avoid obstacle.” Split by map and paraphrase so evaluation instructions are unseen.
04Language-conditioned controller(120 minutes)
Condition the controller on an instruction embedding and train by behavior cloning, optionally retaining the original control objective. Save checkpoints and learning curves.
05Sparse planner(60 minutes)
Implement a rule-based planner first, then optionally a VLM planner, that emits instructions every K steps. Test K in {1, 5, 10, 20}, delayed instructions, and one contradictory instruction.
06Evaluation(60 minutes)
Across at least five seeds compare task success, return, planner calls, end-to-end latency, instruction compliance, and safety violations. Success requires comparable success with materially fewer planner calls. Falsification occurs if direct control matches the system at equal compute or if sparse instructions cause brittle recovery.
07Cleanup(10 minutes)
Remove large checkpoints and videos after extracting aggregate metrics; retain configs, seeds, and result tables.
Controls, failure modes, and safety
Controlled variables
environment seeds, controller capacity, observation stream, training steps, planner compute, and total wall-clock budget.
Failure modes
synthetic-label leakage, instruction ambiguity, controller model error, simulator overfitting, and unequal inference budgets.
Safety
simulation only; do not deploy the learned controller on physical hardware without an independent safety layer and emergency stop.