RoboHacks 2026: Winners and Best Physical AI Projects
A voice command is easy for an AI to understand. Turning it into safe movement through an unfamiliar room is much harder.
RoboHacks brought 120 builders and 25 MARS robots to the YC Office in San Francisco on April 11–12, 2026. Over one weekend, teams connected language and vision models to machines that could navigate, remember objects, handle tools, inspect properties, paint portraits, and respond to dangerous situations.
Zero-Shot Cross-Domain Spatial Memory for MARS took first place overall with a pharmacy robot that learned its surroundings from a single video walkthrough. Other teams built a sarcastic robot chef, an autonomous bomb-disposal unit, a companion for older adults, and a robotic portrait painter.
These are the winning and standout physical AI projects from RoboHacks 2026.
RoboHacks 2026 at a Glance
- Event: RoboHacks 2026
- Dates: April 11–12, 2026
- Location: YC Office, San Francisco
- Participants: 120 builders
- Hardware: 25 MARS robots from Innate
- Overall winner: Zero-Shot Cross-Domain Spatial Memory for MARS by Team DimiMax
- Featured challenges: Google DeepMind’s Best Multimodal Model Use and Scale AI’s Best Use of Data
- Main technologies: Vision language models, spatial memory, SLAM, LiDAR, voice AI, robotic manipulation, and autonomous navigation
- Featured projects: Gordy, RECON, Cloudy, BnBot, Leo, Luma, Rocky, and RedPlanet Cleaning Station
- Explore the event: View the RoboHacks projects and submissions on Iterate
What Was RoboHacks 2026?
RoboHacks was a two-day physical AI hackathon held at the YC Office in San Francisco. The event gave 120 builders access to 25 MARS robots from Innate, allowing teams to test their ideas on shared robotic hardware.
Projects had to connect software intelligence with real-world movement. Teams worked with live camera feeds, language and vision models, LiDAR, SLAM, speech systems, robotic arms, and custom hardware. The resulting robots could perceive their surroundings, follow spoken instructions, navigate rooms, interact with objects, and store information about what they had seen.
Google DeepMind ran the Best Multimodal Model Use challenge, which received 15 submissions. Most entries used Gemini or Gemma to interpret visual information and translate it into robotic actions.
Scale AI ran the Best Use of Data challenge. Its six submissions largely focused on persistent memory, spatial mapping, and the data required for a robot to reason about an environment after moving through it.
RoboHacks 2026 Winner
Zero-Shot Cross-Domain Spatial Memory for MARS.
Team DimiMax took first place overall with an autonomous pharmacy robot that could learn a new environment from a single video walkthrough. It required no training specific to the building in which it operated.
The system reconstructed a 3D scene from monocular video using Pi3. It then aligned that reconstruction with LiDAR occupancy grids through RANSAC and ICP. Local vision language models labelled the resulting environment to create a semantic map, with Gemma 3 handling the labelling on-device.
A GPT-4o Realtime skill orchestrator received voice commands and assigned the required navigation and manipulation tasks. An in-house ACT policy controlled grasping. Together, these components allowed the robot to complete requests suchодаря as fetching a prescription from start to finish.
The project combined Pi3, DUSt3R, VGGT, Gemma 3, an ACT policy, ROS 2, 2D LiDAR SLAM, and the MARS platform. Its spatial map grounded objects visually, semantically, and physically, helping the robot operate in a location it had never encountered before.
Standout Multimodal Robotics Projects
The Google DeepMind challenge asked teams to put a multimodal model inside a physical control loop. The resulting projects used camera feeds, language, and speech to help robots understand instructions and respond to changes in their surroundings.
Gordy: The Physical AI Sous Chef
Gordy was a robot chef inspired by Gordon Ramsay. Built by Team rm -rf /legs on the Innate OS platform, it could listen to an order, speak to the user, observe its workspace, and prepare food.
Gemini handled natural language understanding and checked visually whether each physical action had worked. ElevenLabs provided transcription and expressive speech, allowing the robot to maintain its sarcastic British character throughout the interaction.
The project received seven public votes and entered the Google DeepMind Best Multimodal Use of VLM track.
RECON: Autonomous Bomb Disposal With VLM Intelligence
Team ReCoN built an explosive ordnance reconnaissance unit around the MARS platform. Gemini Flash 2.5 analyzed the scene at 2Hz to identify potential threats and civilians, while SLAM mapping and LiDAR obstacle avoidance supported navigation.
The robot could issue spoken evacuation warnings through ElevenLabs and approach suspicious devices with a three-step P-controller. An arm-mounted camera provided a closer view for wire-level defusal analysis.
A browser dashboard displayed video, radar motion data, and the live SLAM map. An operator could override the autonomous system through natural language commands, while LiDAR safety stops, velocity limits, and command duration limits restricted the robot’s movements.
RECON received seven public votes, the highest total in the Scale AI challenge.
Cloudy: A Companion Robot With Spatial Memory
Cloudy was designed to help older adults living alone find misplaced belongings. A user could ask where an object had been left, and the robot could answer using its record of previous observations.
Gemini 3 Flash Preview converted the live camera feed into structured scene descriptions. These descriptions covered object categories, colors, spatial relationships, and the surfaces supporting each item. Gemini 2.5 Flash interpreted questions and queried the robot’s memory through tool calls.
That memory ran on PostgreSQL and pgvector. It stored 768-dimensional embeddings alongside each object’s 3D position, room, support surface, and temporal history. An association engine compared new observations with existing records using appearance, location, and semantic similarity, reducing duplicate entries.
When the robot had not seen an object for five minutes, it marked the item as missing. This allowed Cloudy to distinguish between an object’s current location and the last place it had appeared.
BnBot: A Robot Concierge for Short-Term Rentals
BnBot handled two recurring tasks for short-term rental hosts: inspecting a property after checkout and welcoming the next guests.
Its inspection agent moved through the property and looked for trash or scratches. Gemini compared images taken before and after the stay, then determined which changes should be reported to the host.
A separate hospitality agent greeted arriving guests and provided a personalized walking tour. The project combined SLAM navigation, scene comparison through a vision language model, speech, an LCD, servo output, and email reporting.
Leo: A Robot That Paints Portraits
Leo turned a conversation and a camera image into a physical portrait. The MARS robot first located a person, spoke with them, and took a photograph using its onboard camera.
Gemini 3.1 converted the image into line art. A computer vision pipeline then extracted centerline paths that the robotic arm could follow to draw the portrait.
The robot could also swap tools through a custom hardware module using NVIDIA GraspGen and Meta’s SAM3. The team organized the system into three layers: agents managed the interaction, skills connected those agents to the painting process, and the Leo pipeline produced the final drawing.
Luma: An Autonomous Lunar Explorer
Team Per Aspera built Luma to classify, collect, and sort lunar regolith inside a lunar-style sandbox. After identifying different rock categories, the robot placed the samples into a custom holder.
A three-agent system divided the workflow between ground control, satellite control, and lunar control. This structure reproduced the project’s command chain, while a vision model classified the rocks and prompted an ACT model to collect and sort them.
Luma received three public votes in the Scale AI Best Use of Data challenge.
Rocky: A Robot Controlled Through Conversation
Rocky was a fully autonomous robot that operated without an app or handheld controller. Users could speak directly to it, while ElevenLabs Conversational AI handled speech recognition, reasoning, and tool calls.
A LangGraph orchestration layer converted requests into sequences of robotic skills. These skills ran through ROS 2 action servers connected with Zenoh.
For navigation, Gemini 2.0 Flash analyzed live camera frames. It determined whether the target appeared in view, how centered it was, how far away it seemed, and which direction the robot needed to move.
The approach sequence followed a fixed loop. Rocky rotated in 15-degree increments, scored each frame, turned toward an off-center object, and moved forward by 20 centimeters once aligned. It released its load after reaching the target. The system ran on NVIDIA Jetson hardware.
RedPlanet Cleaning Station: A Low-Cost Self-Cleaning Robot
Team RedPlanet built a self-reflective agent capable of cleaning itself. The team assembled the project entirely from materials purchased at Target for under $20.
RedPlanet Cleaning Station received three public votes in the Scale AI Best Use of Data challenge.
Technologies Behind the RoboHacks Projects
RoboHacks teams combined several systems rather than relying on one model to control an entire robot. Each layer handled a different part of the physical workflow.
Vision Language Models
→ Gemini and Gemma helped robots interpret live images and connect visual information with language. Teams used them to recognize objects, describe scenes, verify completed actions, compare property images, and estimate the position of a target.
SLAM and LiDAR
→ SLAM allowed robots to map an environment while moving through it. LiDAR supplied distance and obstacle data that supported navigation and safety.
RECON used both technologies to navigate hazardous scenes, while the winning DimiMax project aligned LiDAR occupancy grids with a 3D reconstruction created from monocular video.
Spatial Memory
→ Cloudy and the winning pharmacy robot stored more than a sequence of camera frames. They connected objects with locations, rooms, surfaces, and previous observations.
This gave the robots information they could query later. A robot could remember where an item had appeared or understand where a requested object sat within a larger map.
Voice Interfaces
→ Voice allowed users to direct robots without a separate control interface. Gordy listened to food orders, Rocky converted spoken requests into sequences of skills, and RECON supported natural-language operator commands.
Several projects also used synthetic speech to respond, provide commentary, or broadcast warnings.
Robotic Manipulation
→ Physical tasks required a separate control layer after the models interpreted the request. ACT policies, computer vision pipelines, and custom hardware translated decisions into grasping, sorting, drawing, and tool-changing movements.
The submissions showed a recurring division of work: AI models interpreted the situation, while robotics systems controlled how the machine moved through it.
Build and Test Physical AI Ideas With Iterate!
RoboHacks gave teams access to shared hardware, a clear deadline, and focused technical challenges. In two days, builders moved beyond concepts and demonstrated robots that could navigate, remember, communicate, and manipulate objects.
Could a focused build event help your team turn an AI or robotics idea into a working prototype? Run a hackathon with Iterate and give participants a structured setting to build, test, and present their solutions.
Conclusion
RoboHacks 2026 showed how multimodal models can move beyond screens and interact with physical environments. The winning project combined spatial reconstruction, semantic mapping, voice commands, and robotic manipulation to operate in a previously unseen location.
Across the other submissions, teams applied similar building blocks to food preparation, emergency response, assisted living, property management, art, exploration, and autonomous navigation. Each project approached a different use case, but all had to solve the same underlying problem: turning AI output into reliable physical action.
Frequently Asked Questions
When and where did RoboHacks 2026 take place?
RoboHacks took place at the YC Office in San Francisco on April 11–12, 2026.
How many people participated in RoboHacks?
The event brought together 120 builders and provided 25 MARS robots from Innate.
Which project won RoboHacks 2026?
Zero-Shot Cross-Domain Spatial Memory for MARS by Team DimiMax took first place overall. The project built an autonomous pharmacy robot that learned a new environment from a single video walkthrough.
What were the main RoboHacks challenges?
Google DeepMind ran the Best Multimodal Model Use challenge, while Scale AI ran the Best Use of Data challenge.
What technologies did RoboHacks teams use?
Teams worked with vision language models, SLAM, LiDAR, spatial databases, voice AI, ROS 2, robotic control policies, computer vision, and custom hardware.