A Robot-Oriented Full-Stack Framework for Long-Horizon Embodied Agents
in Evolving Physical and Social Environments

Yicheng Xu¹  ·  Zhikai Xu¹  ·  Ziyuan Chen²  ·  Yuhang Dong¹  ·  Tao Zhang¹  ·  Zhucun Xue¹  ·  Yong Liu¹  ·  Jiangning Zhang¹

¹ Zhejiang University    ² Fudan University

Overview  ·  TODO  ·  Install  ·  Docs  ·  Tasks  ·  Credits  ·  中文

_{Socially grounded interaction · Policy-driven grasping · Evolving tasks · Physical execution}

Connect multimodal reasoning to physical robot execution, embodied human interaction, and changing task requirements.

Infra Harness
Data Curation
Benchmark Evaluation
Real-Robot

Single RTX 4090 · Ubuntu 22.04 / 24.04 · Isaac Sim 5.1 · Python 3.11

The default deployment uses one GPU and one Isaac worker. Agent and support language models run through remote OpenAI-compatible APIs; perception, navigation, and grasp services run locally.

1. Prepare. Install Isaac Sim and the assets and checkpoints, then fill the host template with your paths and model endpoints. Follow the installation guide for Conda, system dependencies, and resource layout.

cp scripts/host_env.example.sh host_env.sh
# Fill local paths and agent/support API credentials in host_env.sh.

2. Install and launch. The default input is a navigation example.

scripts/setup.sh
scripts/evaluate.sh --plan
scripts/evaluate.sh

3. Manage the run or supply your own tasks.

scripts/evaluate.sh --status
scripts/evaluate.sh --stop
scripts/evaluate.sh --episodes /path/to/episodes

--episodes accepts YAML, a directory, or a JSON path index. --plan previews the launch configuration. Runtime files go to ~/.local/share/tarsphere/runs; use --output-dir to choose another directory. Public execution uses the default mechanisms; ordinary history, action, camera, and mapping parameters remain configurable.

Navigate-and-report and fetch-and-deliver share a five-stage causal evaluation. The source examples cover six task conditions:

Evaluation includes Success, Progress, path efficiency, safety, and inference cost. Provide your task collection through --episodes.

tarsphere/
  simulator/       Simulation core and extensions
  embodiment/      Robot profiles and controllers
  infra/           Physical execution and simulator integration
  harness/         Model interface, memory, and skills
  bench/           Episode schemas and task authoring
  evaluation/      Task-state evaluation and scoring
scripts/           Setup, execution, scoring, and rendering
episodes/          Task and condition examples
examples/          Small runnable examples
third_party/       Bundled upstream source and notices

Assets and checkpoints are installed separately. Keep credentials and generated runtime files outside the distributed source.

We thank the authors and maintainers of the works that support TARSphere:

• GRScenes for indoor scene assets.
• OpenFrontier for the visual frontier exploration approach, FrontierNet for frontier prediction, and NavDP / X-NavDP and InternNav for navigation components.
• SAM3, GraspGenX, and cuRobo for perception, grasp generation, and motion planning.
• NVIDIA Isaac Sim for the underlying simulation engine.

Component-specific upstream links, citations, licenses, and local adaptations are documented in third_party. Scene packages, robot assets, Isaac Sim, native character animations, and model weights retain their own terms and are installed separately; see assets.

Copyright 2026 TARSphere contributors. Original project code and documentation are licensed under Apache-2.0.