HARBORRobot RL, as a request
Point it at a simulator. Describe a task. Get a trained policy — with an executable check at every step.
Skip to content
One prompt, End-to-end workflow
Set up simulation · Write the task · Design the reward · Wire the algorithms · Train the policy
One harness, different tasks · robots · simulators
Why a harness
Reinforcement learning works. The pipeline around it is what costs weeks.
Practitioners build the task, shape the reward, calibrate randomization, tune hyperparameters — and pay that cost again for every new simulator, task, and algorithm. Prior automation targets one stage at a time, so the integration burden between stages is repeatedly re-paid by hand.
HARBOR treats the whole problem as harness engineering: shifting human effort from executing each step to designing an agent-readable workflow with verifiable interfaces. Robot RL is unusually well suited to this, because the MDP already exposes stable interfaces — state, action, reward, dynamics, termination — and simulators already produce executable feedback.
HARBOR cannot prove your policy is semantically correct. What it does is turn the common RL engineering failures into gate failures that surface before they propagate downstream.
Long-horizon automation
A reliable workflow over tightly coupled decisions — related work automates individual steps.
Wall-clock efficiency
The iterative stages — reward engineering and hyperparameter tuning — are where the clock goes.
Self-improvement
Human heuristics and existing examples, carried into the next run by in-context learning.
Interpretability and controllability
Full trace documentation at every stage, and human intervention wherever you want it.