MIKASA-Robo-VLA

A memory-intensive robotic manipulation benchmark for Vision-Language-Action research.

Shell Game Touch	Shell Game Shuffle Color Lamp Touch	Remember Color	Remember Shape & Color
Find Imposter Color	Intercept	Rotate Strict	Take It Back

What is MIKASA-Robo-VLA?

MIKASA-Robo-VLA extends the MIKASA-Robo memory benchmark to language-conditioned Vision-Language-Action research. It provides tabletop robotic manipulation environments that require an agent to retain and use information across delayed, occluded, temporal, or multi-stage interactions.

The canonical VLA benchmark contains 90 tasks with natural-language instructions, ManiSkill/Gymnasium environments, and released trajectory datasets for training and evaluation. The benchmark task manifest is mikasa_robo_vla_envs.csv.

What changed from MIKASA-Robo (RL release)

• Task set grows from 32 → 90 registered environments covering 10 memory types (vs 4 in the RL release).
• Every task ships a natural-language LANGUAGE_INSTRUCTION for VLA conditioning.
• Episodes are grouped into three horizon splits (Short / Medium / Long) so multi-task training and evaluation are tractable.
• 22,500 PPO / motion-planning oracle trajectories are released on Hugging Face in RLDS and LeRobotDataset v3 formats — no further conversion needed (6+ million transitions).
• Dense and normalised-dense rewards are calibrated for every task, enabling both offline imitation learning and online RL.
• The original 32-task RL implementation is available from the mikasa-robo-rl branch and remains under mikasa_robo_suite/rl/ for backwards compatibility.

Installation

Install from the repository with the locked uv environment:

git clone https://github.com/CognitiveAISystems/MIKASA-Robo.git
cd MIKASA-Robo
uv sync --frozen

See the installation guide for system requirements, package-install alternatives, and setup troubleshooting.

Quick Start

Every benchmark environment should be wrapped with apply_mikasa_vla_wrappers immediately after gym.make so its observations and task logic match the released VLA data pipeline.

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(
    "RememberColor3-VLA-v0",
    num_envs=1,
    obs_mode="rgb",
    control_mode="pd_ee_delta_pose",
    reward_mode="normalized_dense",
    render_mode="all",
    sim_backend="gpu",
)
env = apply_mikasa_vla_wrappers(env, include_overlays=False)

obs, info = env.reset(seed=42)
for _ in range(env.max_episode_steps):
    action = torch.as_tensor(env.action_space.sample(), device=env.unwrapped.device)
    obs, reward, terminated, truncated, info = env.step(action)
    if torch.as_tensor(terminated | truncated).any():
        break

env.close()

For task browsing, wrapper behavior, language instructions, and the observation/action contract, use the quick start, environment catalogue, and observation/action reference.

Benchmarking

Run the reference checkpoint-free dummy policy first to smoke-test the evaluation pipeline:

uv run python examples/eval_demo.py \
  --num-episodes 1 --sim-backend gpu \
  --output-dir eval_results/dummy

Canonical evaluation is organized by horizon split and uses the benchmark protocol for task selection, seeds, metrics, and result files. See Benchmarking and the Evaluation Protocol before reporting results.

Datasets

MIKASA-Robo-VLA provides the full 90-task trajectory release on Hugging Face. The data pipeline supports:

• NPZ source episodes for local collection and custom preprocessing.
• RLDS / TFDS for episodic dataset pipelines.
• LeRobotDataset v3 for modern PyTorch and VLA fine-tuning workflows.

Download one LeRobotDataset task with huggingface_hub:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="mikasa-robo/mikasa-robo-vla-lerobot",
    repo_type="dataset",
    allow_patterns="remember_color_3_vla_v0/**",
    local_dir="data_mikasa_robo/data_lerobot",
)

allow_patterns matches paths inside the Hugging Face dataset repository. LeRobot task directories use normalized lowercase dataset names, for example RememberColor3-VLA-v0 is stored as remember_color_3_vla_v0/. The downloaded files are placed under data_mikasa_robo/data_lerobot/remember_color_3_vla_v0/.

The dataset guide covers the public RLDS and LeRobot releases, local collection, dataset fields, and export workflows.

Useful Links

• Documentation
• Installation
• Quick Start
• Environments and Tasks
• Benchmarking
• Datasets

Citation

If you use MIKASA-Robo-VLA in your research, please cite:

@inproceedings{cherepanov2026memory,
  title     = {Memory, Benchmark \& Robots: A Benchmark for Solving Complex Tasks with Reinforcement Learning},
  author    = {Egor Cherepanov and Nikita Kachaev and Alexey Kovalev and Aleksandr I. Panov},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year      = {2026},
  url       = {https://openreview.net/forum?id=9cLPurIZMj}
}