eurekaverse-go2-parkour-seed1

Eurekaverse Go2 Parkour Policy — Seed 1

A Unitree Go2 quadruped parkour locomotion policy trained with an LLM-generated environment curriculum (Eurekaverse-style), reproduced on Isaac Lab 3.0.

This is the final policy of seed 1 of a 3-seed reproduction run.

⚠️ This is the privileged teacher policy (it consumes terrain scandots — a height scan around the robot). It is NOT deployable standalone: without scandots it just collapses. To run on a real robot / another sim, distill a depth-based student first (train.py --use_camera ...). See the code repo for setup.

What this is

• Task: blind-ish parkour locomotion — traverse obstacle courses toward 8 sequential goals (ramps, boxes, stepping stones, stairs, gaps, beams, …).
• Training: 5 curriculum iterations × 8 parallel policy runs × 2000 PPO steps/env, resumed from a 1000-step flat-ground walk pretrain. Terrains generated each iteration by gpt-4o-2024-08-06. This checkpoint is the iteration-4 winner (parallel run 1).
• Checkpoint: model_11000.pt (final). model_0.pt is the iteration-4 starting point.

Benchmark performance

Held-out benchmark of 20 parkour tasks × 10 difficulty levels (200 instances), metric = number of goals reached (out of 8):

• This policy: 4.36 / 8
• Distribution across the run's 8 final parallel policies: 4.44 ± 0.34 (range 3.68–4.79)

Per-task numbers are in benchmark_results.txt. Strongest on ramps / pillar-jumps / balance beam (>6/8); weakest on narrow tunnels and spherical speed-bump climbs (<2/8).

Files

How to load / use

This is not a standalone / transformers model. It requires the training codebase and simulator:

• Isaac Lab 3.0 (Isaac Sim) + the extreme-parkour / legged_gym environment used for training.
• Place model_11000.pt and legged_robot_config.pkl under logs/<proj>/<exptid>/, then load with the repo's task_registry (get_saved_cfgs reads legged_robot_config.pkl; the runner loads the .pt). See the project's scripts/evaluate.py.

Note: this is a reproduction/fork ported to Isaac Lab 3.0, not the original Eurekaverse (Isaac Gym) code, so absolute numbers may differ from the paper.

Attribution

• Method: Eurekaverse: Environment Curriculum Generation via Large Language Models (Liang et al., CoRL 2024) — https://eureka-research.github.io/eurekaverse/
• Locomotion backbone: extreme-parkour.
• Trained by the repository owner as an independent reproduction. Verify upstream licenses before redistribution.