mikasa-robo-suite 0.0.2

Gym-like memory-intensive environmtnts for robotic tabletop manipulation

MIKASA-Robo

Benchmark for robotic tabletop manipulation memory-intensive tasks

---

Example tasks from the MIKASA-Robo benchmark

Overview

MIKASA-Robo is a comprehensive benchmark suite for memory-intensive robotic manipulation tasks, part of the MIKASA (Memory-Intensive Skills Assessment Suite for Agents) framework. It features:

• 12 distinct task types with varying difficulty levels
• 32 total tasks covering different memory aspects
• First benchmark specifically designed for testing agent memory in robotic manipulation

Key Features

• Diverse Memory Testing: Covers four fundamental memory types:

  • Object Memory
  • Spatial Memory
  • Sequential Memory
  • Memory Capacity

• Built on ManiSkill3: Leverages the powerful ManiSkill3 framework, providing:

  • GPU parallelization
  • User-friendly interface
  • Customizable environments

List of Tasks

Preview	Memory Task	Mode	Brief Description	T	Memory Task Type
	ShellGame[Mode]-v0	Touch Push Pick	Memorize the position of the ball after some time being covered by the cups and then interact with the cup the ball is under.	90	Object
	Intercept[Mode]-v0	Slow Medium Fast	Memorize the positions of the rolling ball, estimate its velocity through those positions, and then aim the ball at the target.	90	Spatial
	InterceptGrab[Mode]-v0	Slow Medium Fast	Memorize the positions of the rolling ball, estimate its velocity through those positions, and then catch the ball with the gripper and lift it up.	90	Spatial
	RotateLenient[Mode]-v0	Pos PosNeg	Memorize the initial position of the peg and rotate it by a given angle.	90	Spatial
	RotateStrict[Mode]-v0	Pos PosNeg	Memorize the initial position of the peg and rotate it to a given angle without shifting its center.	90	Object
	TakeItBack-v0	---	Memorize the initial position of the cube, move it to the target region, and then return it to its initial position.	180	Spatial
	RememberColor[Mode]-v0	3/5/9	Memorize the color of the cube and choose among other colors.	60	Object
	RememberShape[Mode]-v0	3/5/9	Memorize the shape of the cube and choose among other shapes.	60	Object
	RememberShapeAndColor[Mode]-v0	3×2/3×3 5×3	Memorize the shape and color of the cube and choose among other shapes and colors.	60	Object
	BunchOfColors[Mode]-v0	3/5/7	Remember the colors of the set of cubes shown simultaneously in the bunch and touch them in any order.	120	Capacity
	SeqOfColors[Mode]-v0	3/5/7	Remember the colors of the set of cubes shown sequentially and then select them in any order.	120	Capacity
	ChainOfColors[Mode]-v0	3/5/7	Remember the colors of the set of cubes shown sequentially and then select them in the same order.	120	Sequential

Total: 32 tabletop robotic manipulation memory-intensive tasks in 12 groups. T - episode timeout.

Quick Start

Installation

git clone [e-mail]:CognitiveAISystems/MIKASA-Robo.git
cd MIKASA-Robo
pip install -r requirements.txt

Basic Usage

import mikasa_robo
from mikasa_robo_suite.utils.wrappers import *

num_envs, seed = 512, 123
# Create the environment via gym.make()
# obs_mode="rgb" for modes "RGB", "RGB+joint", "RGB+oracle" etc.
# obs_mode="state" for mode "state"
env = gym.make("RememberColor9-v0", num_envs=num_envs,
                obs_mode="rgb", render_mode="all")

env = StateOnlyTensorToDictWrapper(env) # [always] gen. obs keys

obs, _ = env.reset(seed)
for i in tqdm(range(89)):
    action = torch.from_numpy(env.action_space.sample())
    obs, reward, terminated, truncated, info = env.step(action)
env.close()

Advanced Usage: Debug Wrappers

import mikasa_robo
from mikasa_robo_suite.utils.wrappers import *
from mani_skill.utils.wrappers import RecordEpisode

num_envs, seed = 512, 123
env = gym.make("RememberColor9-v0", num_envs=num_envs,
obs_mode="rgb", render_mode="all")

env = StateOnlyTensorToDictWrapper(env) # [always] gen. obs keys
env = RememberColorInfoWrapper(env) # [debug] show task info
env = RenderStepInfoWrapper(env) # [debug] show env step
env = RenderRewardInfoWrapper(env) # [debug] show total reward
env = DebugRewardWrapper(env) # [debug] show reward info
env = RecordEpisode(env, "./videos/demo_remember-color-9")

obs, _ = env.reset(seed)
for i in tqdm(range(89)):
    action = torch.from_numpy(env.action_space.sample())
    obs, reward, terminated, truncated, info = env.step(action)
env.close()

Video("./videos/demo_remember-color-9/0.mp4", embed=True)

Training

MIKASA-Robo supports multiple training configurations:

PPO with MLP (State-Based)

python3 baselines/ppo/ppo_memtasks.py \
    --env_id=RememberColor9-v0 \
    --exp-name=remember-color-9-v0 \
    --num-steps=60 \
    --num_eval_steps=180 \
    --include-state

PPO with MLP (RGB + Joint)

python3 baselines/ppo/ppo_memtasks.py \
    --env_id=RememberColor9-v0 \
    --exp-name=remember-color-9-v0 \
    --num-steps=60 \
    --num_eval_steps=180 \
    --include-rgb \
    --include-joints

PPO with LSTM (RGB + Joint)

python3 baselines/ppo/ppo_memtasks_lstm.py \
    --env_id=RememberColor9-v0 \
    --exp-name=remember-color-9-v0 \
    --num-steps=60 \
    --num_eval_steps=180 \
    --include-rgb \
    --include-joints

To train with sparse rewards, add --reward-mode=sparse.

MIKASA-Robo Ideology

The agent's memory capabilities can be accessed not only when the environment demands memory, but also when the observations are provided in the correct format. Currently, we have implemented several training modes:

• state: In this mode, the agent receives comprehensive, vectorized information about the environment, joints, and TCP pose, along with oracle data that is essential for solving memory-intensive tasks. When trained in this way, the agent addresses the MDP problem and does not require memory.

• RGB+joints: Here, the agent receives image data from a camera mounted above and from the manipulator's gripper, along with the position and velocity of its joints. This mode provides no additional information, meaning the agent must learn to store and utilize oracle data. It is designed to test the agent's memory capabilities.

These training modes are obtained by using correct flags. Thus,

# To train in `state` mode:
--include-state

# To train in `RGB+joints` mode:
--include-rgb \
--include-joints

# Additionally, for debugging you can add oracle information to the observation:
--include-oracle

Collecting datasets for Offline RL

1. Run training PPO-MLP on MIKASA-Robo tasks in the state mode (i.e. in MDP mode with oracle information):

# For single task:
python3 dataset_collectors/get_dataset_collectors_ckpt.py --env_id=ShellGameTouch-v0

# For all tasks:
python3dataset_collectors/parallel_training_manager.py

2. Collect datasets using oracle checkpoints:

# For single task:
python3 dataset_collectors/get_mikasa_robo_datasets.py --env-id=ShellGameTouch-v0 --path-to-save-data="data" --ckpt-dir="."

# For all tasks:
python3 dataset_collectors/parallel_dataset_collection_manager.py --path-to-save-data="data" --ckpt-dir="."

Citation

If you find our work useful, please cite our paper:

@misc{cherepanov2025mikasa,
      title={Memory, Benchmark & Robots: A Benchmark for Solving Complex Tasks with Reinforcement Learning}, 
      author={Egor Cherepanov and Nikita Kachaev and Alexey K. Kovalev and Aleksandr I. Panov},
      year={2025},
      eprint={2502.10550},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2502.10550}, 
}