How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
Copy for public landing · GitHub Pages or standalone site
Headline
What changed between these two robot runs?
Subhead
Mismatch Lab helps teams understand robot behavior — and spot when a difference means the model no longer fits reality.
Primary CTA: Try Robot Diff
Secondary CTAs: Explore cases · Run the benchmark · Read the research
Preloaded case: Persistent target drift
Run A (baseline policy) Run B (after drift onset)
↓ ↓
Prediction track Observation track
↓ ↓
Difference grows at t=8 … t=35
↓
Repair ×3 did not absorb error
↓
Adequacy hint: consider structural review
User does not read a paper to understand the problem.
| Verb | One line |
|---|---|
| Replay | Watch robot rollouts like video — with prediction overlay |
| Diff | Compare two runs; auto-find behavior shifts and divergence |
| Explain | Timeline of interesting moments and what the model expected |
| Discover | Search thousands of rollouts for episodes worth inspecting |
Fine print (tooltip): Adequacy insights appear when differences persist after recovery attempts.
Headline: Upload a rollout. Get a diff report.
If upload not ready:
Join the pilot — early access to Model Adequacy SDK and batch diff on your logs.
Form fields: name · email · org · stack (Isaac / custom / sim-only) · use case
| Case | User question | Adequacy? |
|---|---|---|
| Policy A vs B | What changed after my ablation? | No |
| Sim vs real | Where does sim diverge from hardware? | Maybe |
| Repair plateau | Why didn't retuning help? | Yes |
| Drift vs noise | Is this sensor noise or real dynamics change? | Yes |
Each card opens interactive Diff viewer with timeline slider.
Headline: Built on open research
Links: Research OS · prereg · benchmark spec
Long term: physical AI that keeps learning when reality breaks its assumptions.
Not the hero. One paragraph max.
Title: Mismatch Lab — Robot Diff & Model Adequacy
Description: Compare robot rollouts, explain behavior shifts, and detect when your world model needs more than retuning.
OG image: Side-by-side diff timeline screenshot
Diff Explorer | Benchmark | Research | API Docs | Join Pilot
Updated 2026-07-29