Skip to content

Crippled Ant — the environment

MuJoCo's Ant-v5 with joints disabled: a quadruped that has to keep walking after losing the use of a leg. It is the reality gap in miniature — the body the policy was trained on is not the body it has to control.

This page is what the environment is. To train and evaluate on it, see Running the Crippled Ant.

Prerequisites

A working rlbootcamp environment. All commands assume it is active and that you are in the package directory:

cd "sessions/02-intermediate/crippled-ant"

Layout

crippled-ant/
├── envs/crippled_ant.py       # CrippledAnt wrapper + make_ant factory
├── scripts/train.py           # Hydra training CLI (PPO / SAC)
├── scripts/evaluate.py        # evaluate a run, optionally under a new injury
├── scripts/render_agent.py    # render any checkpoint .zip to .mp4, any injury
├── scripts/transfer_benchmark.py   # injury-severity benchmark (Ex 3/4)
├── conf/                      # config.yaml, algo/{ppo,sac}.yaml, experiment/
└── tests/                     # contract + smoke tests

The wrapper

import sys; sys.path.insert(0, '.')
from envs import CrippledAnt, make_ant
import gymnasium as gym

# explicit joints…
env = CrippledAnt(gym.make("Ant-v5"), disabled_joints=[2, 3])
# …whole legs (a leg = hip + ankle action pair)…
env = make_ant(disabled_legs=[0])
# …or domain randomisation: new random legs every reset, fixed count
env = make_ant(n_random_legs=1)
# …or randomise the SEVERITY too: count sampled uniformly on every reset
env = make_ant(n_random_legs_max=4)   # 0 is in the support -- a healthy Ant
                                      # is part of the training distribution

obs, info = env.reset(seed=0)
info["disabled_joints"]          # verify what is actually disabled
info["n_disabled_legs"]          # how many legs — 0 under n_random_legs_max

Ant-v5 action layout — one (hip, ankle) pair per leg:

leg joints position
0 0, 1 front left
1 2, 3 front right
2 4, 5 back left
3 6, 7 back right

The wrapper zeroes torques, not observations: the policy still senses the dead leg, it just cannot move it. Injuries are reproducible per seed, and the active injury is reported in info["disabled_joints"] on every reset and step.

The contract

pytest                # wrapper + config contracts (< 1 s)
pytest -m slow        # end-to-end train/evaluate/benchmark smoke (~2 min)

The suite defines the contract any modified injury implementation must keep: torque zeroing, no caller-side action mutation, unchanged spaces, seed reproducibility, and honest info reporting. If you change the wrapper in Session #2, this is what tells you whether you broke it.

Next