Take It Back

Language Instruction

Push the cube onto the red target, and when the target changes color, return the cube to its original position.

Task Description

Spatial restoration: push a cube to a target, then return it to its original position after the target changes.

Source

• Module: mikasa_robo_suite.vla.memory_envs.take_it_back_vla

• Source file: mikasa_robo_suite/vla/memory_envs/take_it_back_vla.py

Difficulty and Parameters

Two-stage spatial memory and manipulation.

Variants

Env ID	Horizon	Data Source
TakeItBack-VLA-v0	60	PPO

Run Example

import gymnasium as gym
import torch

import mikasa_robo_suite.vla.memory_envs  # registers VLA env IDs
from mikasa_robo_suite.vla.utils.apply_wrappers import apply_mikasa_vla_wrappers

env = gym.make(
    "TakeItBack-VLA-v0",
    num_envs=1,
    obs_mode="rgb",
    control_mode="pd_ee_delta_pose",
    render_mode="all",
)
env = apply_mikasa_vla_wrappers(env)
obs, info = env.reset(seed=42)
for _ in range(env.max_episode_steps):
    action = env.action_space.sample()
    if not torch.is_tensor(action):
        action = torch.as_tensor(action, device=env.unwrapped.device)
    obs, reward, terminated, truncated, info = env.step(action)
env.close()

Dataset Collection

PPO-sourced MIKASA-Robo-90 variants use oracle checkpoints:

uv run python mikasa_robo_suite/vla/dataset_collectors/get_mikasa_robo_datasets.py \
  --env-id TakeItBack-VLA-v0 \
  --path-to-save-data data_mikasa_robo \
  --ckpt-dir . \
  --num-train-data 250

Render Videos

Generate a fresh MP4/GIF render with:

uv run python utils/prepare_benchmark_demo_videos.py \
  --tasks TakeItBack-VLA-v0 \
  --output-dir videos/benchmark_demos \
  --max-attempts-per-task 8 \
  --overwrite

Generated media stays under videos/ and should be published deliberately rather than committed by default.