Stanford HomeBody: Humanoids Gain Autonomous Exploration and Persistent Spatial Memory Across Domestic Environments

The Three-Question Framework

What Happened?

The Stanford Movement Lab (TML) unveiled 'HomeBody', an embodied AI framework that equips bipedal humanoid robots (demonstrated on the Unitree G1 platform) with open-ended spatial exploration, persistent 3D episodic memory, and autonomous task execution in unmapped human households. Without prior CAD models or teleoperated walkthroughs, the humanoid autonomously patrols rooms, builds dynamic semantic point clouds, and catalogs object locations over time.

Why It Matters

Most humanoid robotics demonstrations to date rely on brittle, pre-programmed trajectories in controlled laboratory testbeds. When placed in real homes where humans constantly move chairs, cups, and keys, conventional robots freeze. HomeBody solves this by coupling frontier Vision-Language Models (VLMs) with a topological scene graph that updates asynchronously. When asked in natural language ('Where did I leave my reading glasses?'), the robot queries its episodic memory, traverses the physical house, and retrieves the object.

What Could Happen Next

The Stanford team is open-sourcing the HomeBody memory stack and planning collaborative trials with consumer robotics manufacturers to integrate long-horizon manipulation primitives onto commercial bipedal platforms.

Executive Key Takeaways

Persistent Memory for Humanoids: First system enabling domestic humanoids to retain physical object locations across multi-week horizons.

Zero Pre-Mapping: Explores novel domestic layouts autonomously using active visual SLAM and VLM frontier exploration.

Semantic Scene Graphs: Indexes dynamic environments into queryable 3D topological nodes rather than flat geometric meshes.

Hardware Validated: Proven on commercial Unitree G1 bipedal robots performing manipulation in real-world multi-room apartments.

In a major leap toward truly useful domestic robotics, researchers at the Stanford Movement Lab (TML) introduced HomeBody, an integrated embodied AI architecture that grants humanoid robots persistent episodic spatial memory and open-ended exploration capabilities.

While robotic manipulation models have advanced rapidly with transformer-based policy networks, long-horizon autonomy in unstructured residential spaces has remained an intractable challenge. Human homes are dynamic: clutter accumulates, doors open and close, and everyday objects migrate continuously between rooms.

HomeBody bridges the gap between high-level reasoning foundation models and real-time physical locomotion. Deployed on a commercial Unitree G1 bipedal humanoid equipped with stereo depth cameras and an onboard compute module, the robot enters an unfamiliar home with zero prior architectural maps.

Using active visual SLAM combined with a frontier-based exploration heuristic guided by a Vision-Language Model (VLM), the humanoid systematically walks through hallways and living spaces. As it traverses the home, it constructs a real-time 3D semantic Gaussian splatting representation overlaid with a topological scene graph.

Crucially, HomeBody treats memory as an evolving, multi-tiered database. High-level spatial landmarks (couches, tables, doorways) form rigid anchor nodes, while movable objects (keys, mugs, laptops, books) are registered as probabilistic child entities stamped with temporal observation vectors.

During extensive empirical trials in multi-room apartments, human occupants asked the robot natural language questions regarding displaced belongings (e.g., 'Where is my work laptop?'). The Unitree G1 queried its episodic scene graph, generated collision-free bipedal footstep trajectories, navigated to the target bedroom, and verified the object's presence with an 88.4% success rate across 14-day continuous evaluations.

By eliminating the requirement for manual teleoperated pre-scanning, HomeBody provides the missing foundational software stack needed to transform humanoid robots from choreographed factory demonstration units into adaptable domestic assistants.

The Three-Question Framework

What Happened?

Why It Matters

What Could Happen Next

Executive Key Takeaways

Primary Sources & Citations