TyGrit — Tying Grit

A Research Platform for Mobile Manipulation in Unknown Environments

TyGrit provides infrastructure to study mobile manipulation under uncertainty — where the robot discovers the world through interaction rather than receiving a complete model upfront. See Why a New Framework for the problem formulation.

Key Features

• Whole-body motion planning — VAMP-based planner with an extensible MotionPlanner protocol

• Neural grasp prediction — GraspGen 6-DOF grasp synthesis with GraspPredictor protocol

• Segmentation — Ground-truth (sim) and SAM 3 backends with Segmenter protocol

• Multi-backend environments — ManiSkill 3 simulation via RobotBase protocol

• Receding-horizon control loop — scheduler with config-driven subgoal generation

• Vendored C++ IK solvers — IKFast (analytical) + TRAC-IK (numerical)

• Visualization toolkit — MomaViz: Blender renders, ManiSkill replays, video

Why a New Framework

Why standard frameworks fail and what makes this problem fundamentally different.

Why a New Framework

Getting Started

Prerequisites, installation, and environment setup.

Setup

Architecture

Hierarchical policy design and module overview.

Architecture

Configuration

All TOML sections and parameters.

Configuration

Visualization

MomaViz: Blender renders, ManiSkill replays, video.

Visualization

RL Baseline

FPPO (CausalMoMa) training and evaluation.

RL Baseline (FPPO)