Procedural, hidden scenes
Agents see one example scene. Grading runs their code on seeded scenes they never saw, so memorized or hard-coded answers score zero.
RL environment datasets · Harbor format
Verifiable environments where frontier models still fail.
We build collections of RL environments, organized by domain. Each task has hidden scenes and a programmatic reward, and ships with baselines showing where today's models break. Our first collection covers the physical world.
physical / box-stacking-stability
tilt 0.0°
center of mass gravity
Collections
One collection per domain
Every collection follows the same format and the same verifier rules, so you can buy one domain or several and run them all with the same harness.
Baselines
Where today's models fail
Each model gets k attempts through the same terminus-2 agent harness. Points on the diagonal are consistent; points below it can solve the task but don't do it reliably. That gap is the signal RL training closes, and it is what we ship each environment with.
Mean reward
Closed frontierOpen-weight
Model outputs
One drawing, a dozen apartments
From Floor plan to 3D scene: each model reads the plan and builds the apartment in Blender, then renders it. We show one typical attempt per model. Pick a model to compare its scene with the drawing, and switch to 3D to orbit the Blender scene it built.
Catalog
Environments
Select the environments you're interested in, then request samples. Tell us which in-development ones you'd want, and we'll prioritize them.
Method
Built to be trained on, not just scored
Isolated verifier
Rewards come from MuJoCo or exact geometry in a separate container. Agent code runs unprivileged and cannot read the grader, seeds or ground truth.
Dense rubrics
Each reward breaks down by skill (pose vs pixel, stability vs height, tracking vs identification), so partial competence earns partial credit and RL has signal.
Calibrated difficulty
Every task ships a score ladder of known-flawed strategies, plus frontier and open-weight baselines. QC checks oracle = 1 and no-op = 0 before release.
Samples
Pick what you want to see
A sample is the agent-facing side of an environment, ready to run in Harbor: the task instruction, task.toml, example scenes and Dockerfiles, plus one full model trajectory. Pick the ones you want and we'll email them to you.
Hidden scenes, verifiers and reference solutions stay with us, so samples can't contaminate evaluation.
We also build custom environments for specific embodiments, simulators and skills.