GenesisBench

How language intelligence can be used to improve physical intelligence.

How it works

GenesisBench evaluates how LLM agents can train the policies that run on real robot hardware.

Heuristic policies in simulation

No training and no H100s yet — agents iterate on code logic alone, optimizing heuristic policies for performance in simulation.

This stage ships nine tasks — each runs the same improvement loop in its own simulated environment.

01 / 09MuJoCo Ant

Improve a programmatic Ant locomotion policy under a fixed autonomous research budget.

simulation_heuristics_ant_v1 ↗
Starter policya runnable CPG/PD gait — weaker than the reference
Submit the best policyhidden nominal + dynamics suites · locomotion return · starter = 0, article MPC reference = 100

The loop nodes illustrate gait and planning edits for the Ant v1 task; all nine task packages live on GitHub.

Environments

Two environments: simulation for training and evaluation, the real world for operation.

Environments

Simulation

Policies are trained and evaluated in simulation — the simulator shows how they are really working.

Real-world operation

The goal: policies that can be used to run on real robotics hardware.

real robot hardware

Results

Leaderboards

Five-trial mean ± sample standard deviation for each environment's native raw score. Fail-closed raw-null trials use the starter-equivalent anchor. Dotted markers show the starter; dashed markers show the article reference.

Loading task leaderboards…

Final normalized score

Pooled IQM across all 45 trial-task scores; ± is the sample standard deviation of five per-trial IQMs. Positive display index = IQM + 100; the dotted marker is the starter-level aggregate.

Loading final leaderboard…

Research background

GenesisBench builds on research at the intersection of agentic coding, post-training, and embodied AI.

ENPIRE: Agentic Robot Policy Self-Improvement in the Real WorldNVIDIA GEAR · arXiv:2606.19980 Xiao, Xie, Zhang, Lin, et al. — LLM agents improving robot policies through closed-loop real-world iteration.
All references11