SpatioTemporal

Congratulations to Heba Khamis and the team at Contactile on winning the Propel-AIR competition. Contactile is building remarkable technology, and this recognition is thoroughly deserved. We look forward to seeing what comes next! SpatioTemporal is also seriously proud to have been named runner-up. We entered Propel-AIR to make a clear case: Robots will need more than perception and planning. They will be operating around people - so they must also understand our movement, anticipate our intent and respond naturally in shared spaces. To have that thesis recognised by such an experienced panel is meaningful validation of both the problem and the direction we are pursuing. Thank you to ARM Hub for creating this opportunity, and to the judges for the time, attention and thoughtful engagement throughout the process. We are also grateful to everyone who supported us, challenged the thinking and helped sharpen the story. We may not be taking home first place, but we leave Propel-AIR with stronger conviction, valuable new relationships and real momentum. Congratulations again to Contactile. A worthy winner, and an exciting Australian Physical AI company. Onwards and upwards!

David Pearce has this exactly right. The incident at CeMAT is not so much an isolated robot failure than as a marker of where humanoid deployment stands today. When one robot made unexpected contact with a person, every humanoid at the event was paused. In the absence of an established protocol, the organisers made the safest decision available. That was responsible. It was also revealing. Physical AI is moving into public and workplace environments faster than the operating frameworks around it are being established. Deployment readiness can’t only mean that a robot is technically capable of performing its task. It must also mean that the organisation knows how to respond when its behaviour becomes uncertain. David is right that these protocols need to be designed before they’re needed. We’d add one further layer. The long-term answer can’t rely entirely on humans managing robot uncertainty from outside the system. Robots themselves need a better ability to interpret the movement of the people around them, anticipate emerging interactions, and recognise when it’s safer to yield, slow down or stop. The industry needs both: - Clearer protocols around the robot. - Better human awareness within it. That combination will determine whether capable machines become trusted participants in shared spaces.

Gesa Schneider has put her finger on something important: a world model might help a robot understand objects, physics and likely outcomes, but that still isn't the whole job. A robot also needs some understanding of itself, the people around it, the task they're sharing, and the interaction that's unfolding between them. Those models can't sit in isolation. They need to keep updating one another as people move, hesitate, respond, adapt and recover. That builds directly on the story we're telling at SpatioTemporal - Perception tells a machine what's there. - Planning tells it what to do. - But the difficult part sits in between: understanding what movement means, what someone is likely to do next, and what might happen if the machine acts. That's why trust can't be added at the end as a feature. It has to be built through repeated, predictable interaction. Gesa's broader point matters too: better technology won't be enough to drive Physical AI adoption. It'll take trusted partnerships, clear commercial value, strong ecosystems and real collaboration between researchers, manufacturers, customers and technology companies. The next generation of Physical AI won't succeed just because it's more capable. It'll succeed because people trust it enough to feel safe around it.

We're excited to be joining the 2026 PropelAIR cohort. Physical AI is reaching an important point. Robots are becoming increasingly capable at seeing the world and planning their actions. The next challenge is helping them understand the people they share that world with. At SpatioTemporal, we're building Motion Intelligence: technology that compresses space and time into motion tokens so machines can interpret movement, anticipate intent, and behave more naturally in our homes and in our workplaces. We believe this is a critical capability for the next generation of robotics and autonomous systems. Over the coming weeks we're looking forward to working alongside an outstanding group of founders, researchers and industry partners, and to challenging our ideas through the PropelAIR program. Thank you to the ARM Hub team, and everyone who has supported us so far. We're looking forward to sharing what comes next! #PropelAIR #PhysicalAI #Robotics #MotionIntelligence #AI #NVIDIA #SpatioTemporal

We’re thrilled to share that SpatioTemporal has been named a Victorian Finalist in the 2026 AIIA iAwards - in the Artificial Intelligence Technology category. It’s encouraging to see recognition for a field that is still emerging: giving machines the ability to understand motion, infer intent, and operate more safely in shared human spaces. Our belief remains simple. The next generation of Physical AI will need more than perception and planning. It will need machines that can read the room, understand the grammar of movement, and anticipate what might happen next. We’ll need to trust our robot friends before we can fully invite them into our homes. A huge thank you to the iAwards judging panel, and to everyone who has supported the journey so far. We’re especially grateful to our friends and supporters at Natural Velocity, including Nadine ‘Noodles’ Groves MAICD and John McGiffin and the team, for their ongoing encouragement and belief in what we’re building. Congratulations as well to all of this year’s finalists. We look forward to joining you at the Victorian iAwards ceremony in Melbourne next month. #iAwards #ArtificialIntelligence #PhysicalAI #Robotics #Autonomy #MotionIntelligence #SpatioTemporal

Thank you to the ROI from SMEC AI team for featuring Andrew Ballard and SpatioTemporal. The point they highlighted matters: it's not so much that physical AI needs larger models... it's that is needs the right models. Our focus is on compressing space and time into motion tokens, so machines can better understand how people, vehicles and robots move. It matters because real-world autonomy is not a perception problem. A robot may see a person. A vehicle may detect another car. But the harder question is what that motion means: Is someone hesitating? Is a driver drifting? Is a pedestrian about to cross? Is a robot creating risk by moving too late, too fast, or too mechanically? These are motion problems. Our thesis is that the missing middle layer between perception and planning is Motion Intelligence: a compact, intent-aware representation of movement that can help machines anticipate what is likely to happen next. One of the encouraging parts of this work is its efficiency. Motion does not need to be represented as full video to be useful. By learning from structured movement rather than raw appearance, we can build smaller, more focused models that are designed for real-world deployment, including phones, edge devices and eventually lightweight robotics platforms. That's not a compromise. It's the goal. For Australia, this is an important lesson. The next generation of AI companies will not all be built by outspending the largest labs. Some will be built by finding the right abstraction, focusing on a hard physical problem, and proving that a smaller model can do something large systems still struggle with. That's the direction we're pursuing at SpatioTemporal: Robots that read the room. Cars that read the road. Watch the full episode here: https://lnkd.in/djsHN_3Y

We have published a new SpatioTemporal research thesis: State Is All You Need - Introducing World State Vectors for Physical AI The central idea is simple: - World Models are the raster layer of Physical AI.  - World State Vectors are the vector layer. Raster world models are powerful. They help machines simulate, generate, reconstruct and reason over rich physical scenes. But real-time autonomy also needs something thinner: a compressed, action-ready representation of what matters now. That is the role of the World State Vector. - It sits between upstream perception and downstream planning. - It doesnt replace sensors, cameras, lidar, simulation or object detection. It takes their outputs as inputs. - It doesnt replace robot-specific planning or control. It hands off a cleaner state to the systems responsible for action. Internally, the SpatioTemporal architecture has three parts: 1. Spatial Intelligence - Understanding what matters now through motion, intent, relevance and concern. 2. World State Vector - Compressing that into an operational state. 3. Temporal Intelligence - Using that state to reason about futures, consequence and planning handoff. The operating principle is: Start with vector. Fall back to raster when detail is required. This is how we see the missing middle layer for Physical AI: a lightweight, embeddable intelligence layer that helps robots and autonomous systems understand motion, infer intent and act from a clearer state of the world. Robots that read the room.  Cars that read the road. #PhysicalAI #Robotics #WorldModels #SpatialComputing #AutonomousSystems https://lnkd.in/gNggc44S

A fighter pilot does not need a perfect copy of the sky. They need only a cockpit telling them what matters now. That is one of the most useful lessons for 𝗣𝗵𝘆𝘀𝗶𝗰𝗮𝗹 𝗔𝗜. Andrew Ballard of SpatioTemporal makes a point this week that I think is underappreciated in Physical AI: the 𝗯𝗼𝘁𝘁𝗹𝗲𝗻𝗲𝗰𝗸 is not how much reality a system can perceive. It is what it can 𝗽𝗿𝗶𝗼𝗿𝗶𝘁𝗶𝘀𝗲, 𝗶𝗻𝗳𝗲𝗿 and 𝗮𝗰𝘁 𝗼𝗻. I wrote recently about the 𝗳𝗶𝗴𝗵𝘁𝗲𝗿 𝗽𝗶𝗹𝗼𝘁 𝗮𝘀 𝗮𝗻 𝗮𝗻𝗰𝗲𝘀𝘁𝗼𝗿 𝗼𝗳 𝘁𝗵𝗲 𝗿𝗲𝗮𝗹-𝘁𝗶𝗺𝗲 𝗼𝗽𝗲𝗿𝗮𝘁𝗶𝗼𝗻𝗮𝗹 𝘁𝘄𝗶𝗻 - and this is precisely why: The cockpit does not give a pilot more data. It gives them what matters, what is changing, what requires action now. The 𝙨𝙞𝙜𝙣𝙖𝙡-𝙣𝙤𝙩-𝙨𝙩𝙤𝙧𝙖𝙜𝙚 insight Ballard develops in his article for robotics and autonomous systems is the same principle, applied at a different layer. There is a harder version of this question that I will come back to next week: 𝘞𝘩𝘦𝘯 𝘮𝘶𝘭𝘵𝘪𝘱𝘭𝘦 𝘢𝘶𝘵𝘰𝘯𝘰𝘮𝘰𝘶𝘴 𝘴𝘺𝘴𝘵𝘦𝘮𝘴 𝘴𝘩𝘢𝘳𝘦 𝘢𝘯 𝘦𝘯𝘷𝘪𝘳𝘰𝘯𝘮𝘦𝘯𝘵, 𝘸𝘩𝘰 𝘨𝘰𝘷𝘦𝘳𝘯𝘴 𝘵𝘩𝘦 𝘰𝘱𝘦𝘳𝘢𝘵𝘪𝘰𝘯𝘢𝘭 𝘵𝘳𝘶𝘵𝘩 𝘵𝘩𝘦𝘺 𝘢𝘤𝘵 𝘰𝘯? Good weekend reading: https://lnkd.in/dbw96p-A My related post on the pilot and the operational twin: https://lnkd.in/dajSum4q #PhysicalAI #DigitalTwins #Robotics #SpatialComputing #AutonomousSystems

In the last 24 hours, we’ve built a new shared-space robotics simulator around Motion Intelligence, directly showing behaviour with/without our Large SpatioTemporal Model. Each robot maintains its own ego-centric understanding of the world: what it’s trying to achieve / what nearby agents are doing / which movements suggest yielding, hesitation, crossing, or conflict. The coloured markers above each robot show what it currently perceives around it: other robots, nearby humans, and unfolding motion in shared space. What’s interesting is not just collision avoidance, but the beginnings of behavioural adaptation: the robots begin to yield earlier, navigate more fluidly, and make smoother decisions around uncertainty because they’re no longer responding only to geometry. They’re responding to motion as intent. Humans learn this instinctively as children. We read movement continuously: confidence, distraction, gaps, hesitations, and danger. It’s one of the most fundamental human skills, yet robotics systems still struggle to model it directly. At SpatioTemporal, we believe motion is a first-class data source. Our Large SpatioTemporal Model compresses space and time into machine-readable motion tokens, allowing robots to better understand behaviour and make better plans to achieve their goals safely around people. Robots that read the room, just like humans do.