Implementation:Haosulab ManiSkill PushT

Knowledge Sources	Haosulab_ManiSkill
Domains	Robotics, Simulation, Tabletop_Manipulation
Last Updated	2026-02-15 08:00 GMT

Overview

Concrete implementation of the Push-T task environment in ManiSkill, where a robot must push a T-shaped block to a target pose.

Description

The PushTEnv places a T-shaped block on a white table and requires a robot with a stick end-effector to push the block to match a goal pose (both position and rotation). The environment uses a custom white table scene builder.

Registered as PushT-v1 with max_episode_steps=100. The supported robot is panda_stick (PandaStick). Reward modes include "normalized_dense", "dense", "sparse", and "none".

The T-block is built from two overlapping box shapes. Randomizations include the T-block's xy position and z-rotation, and the goal T pose. The goal is visualized with a green outline. Success is evaluated when the T-block position is within 0.01m and rotation is within a threshold of the goal pose.

Usage

Use this environment for non-prehensile manipulation research. The Push-T task is a well-known benchmark for testing pushing and planar manipulation policies, particularly with diffusion-based and behavior cloning approaches.

Code Reference

Source Location

• Repository: Haosulab_ManiSkill
• File: mani_skill/envs/tasks/tabletop/push_t.py

Signature

@register_env("PushT-v1", max_episode_steps=100)
class PushTEnv(BaseEnv):
    SUPPORTED_ROBOTS: ["panda_stick"]
    agent: PandaStick

Import

import gymnasium as gym
import mani_skill.envs
env = gym.make("PushT-v1")

I/O Contract

Inputs

Name	Type	Required	Description
obs_mode	str	No	Observation mode
reward_mode	str	No	Reward mode: "normalized_dense", "dense", "sparse", "none"
control_mode	str	No	Control mode for panda_stick robot

Outputs

Name	Type	Description
obs	dict/array	Observation including T-block pose, goal pose, TCP pose
reward	float	Reward based on position and rotation alignment with goal
terminated	bool	Whether episode ended by success/failure
truncated	bool	Whether episode hit max steps (100)
info	dict	Contains success flag, position error, rotation error

Usage Examples

Basic Usage

import gymnasium as gym
import mani_skill.envs

env = gym.make("PushT-v1", obs_mode="state", render_mode="rgb_array")
obs, info = env.reset()
for _ in range(100):
    action = env.action_space.sample()
    obs, reward, terminated, truncated, info = env.step(action)
    if terminated or truncated:
        obs, info = env.reset()
env.close()

Related Pages

• Environment:Haosulab_ManiSkill_Python_SAPIEN_Core