Robotics Briefing /
Learning from touch and knowing when to give way.
Robots need to respond to information that changes as they move: pressure at a fingertip, a person blocking the camera, or a pull on an arm. This edition examines three approaches to those problems, with evidence ranging from controlled physical trials to simulation and demonstrations. In construction, Origin is taking a related challenge, maintaining useful contact with a wall, into customer deployments.
Your quick takeaways
- Spark-0 uses a thousand hours of human touch data during pretraining. On five physical manipulation tasks, success reached 52%, but no method completed the full test-tube insertion task.
- VP-EVT moves to preserve its view of the person it follows. Simulated tracking success improved substantially, while collision-free performance remained a constraint.
- Compliant GMT lets a humanoid yield to contact while following varied motions. Physical demonstrations are promising; choosing the right stiffness automatically remains future work.
Papers
Spark-0: touch can teach a robot what to look for
Tianxing Chen and colleagues, Xspark AI and university collaborators • October 15 preprint • Catch-up
A camera can show fingers around an object while missing whether they are slipping or pressing too hard. Spark-0 uses touch both to teach visual representations about contact and to guide actions when tactile sensors are available.
Its HumanTouch collection records a thousand hours of human manipulation, using gloves whose sensor layout also fits the robot hands. Contact maps supervise where the visual model should attend during training; a separate pathway supplies the action model with changing pressure information.
Across five physical tasks, with ten trials per task, Spark-0 achieved 52% success versus 36% for the strongest tested baseline, FTP-1. A matched variant that discarded touch during pretraining reached 44%. Both variants used touch during subsequent fine-tuning and physical execution, so that comparison measures the added value of tactile pretraining.
The results support collecting human contact data alongside video. They also show how much remains unsolved: Spark-0 completed none of the full three-slot test-tube insertion trials, despite making partial progress. The physical evaluation uses one robot configuration and only fifty trials per method.
VP-EVT: move before the person being followed disappears
Yanci Wen and colleagues, Southern University of Science and Technology and Tencent Robotics X • October 15 preprint • Catch-up
Following at the right distance can still leave a robot staring at the wrong person’s back. VP-EVT treats keeping the intended person visible as an explicit goal when choosing how to move.
The system combines language-guided visual identification, recent camera history and the robot’s own movement to estimate where the target and nearby pedestrians are. It predicts whether continuing along the current route will prolong an obstruction, then intervenes to change viewpoint when needed.
On two hundred simulated episodes designed around persistent obstruction, tracking success reached 77.5%, versus 35% for OA-VAT. The average longest uninterrupted loss of the target fell from 13.432 to 3.442 seconds. On the separate ordinary-following set, success matched its underlying follower at 90%.
The tradeoff is important: 90.5% of difficult episodes were collision-free, below the underlying follower’s 97.5%. That follower rarely completed the difficult tracking task, but its lower collision rate remains a useful comparison. A physical Unitree Go2 demonstration shows the maneuver working with off-board computing; it does not establish the simulation success rate in crowds.
Compliant GMT: a humanoid can follow a motion and still yield
Venki Parsuram, Michelle Wang, Jin-Hyuk Hong and Pulkit Agrawal • October 15 preprint • Catch-up
A humanoid trained to imitate motion precisely may resist when someone holds its arm. Compliant GMT adds an adapter that adjusts the requested motion so the robot can give way while maintaining balance.
The underlying motion tracker is first trained to tolerate external forces. The adapter then learns from a simulated teacher how to modify the reference motion at different stiffness settings. At deployment, it uses recent measurements of the robot’s own body state rather than requiring the teacher’s direct knowledge of external force.
Physical Unitree G1 examples include yielding during a dance and human-guided interactions such as opening a door and wiping a surface. The simulated tracking evaluation also exposes a compromise: at one stiffness setting, the robot stayed upright in 88% of dynamic-motion windows, compared with 90.4% for SONIC. That metric measures remaining upright during the tested motions, not task completion.
The practical contribution is adjustable yielding across varied movements. Physical manipulation examples are teleoperated, and the work does not establish autonomous household reliability. Extending responses to contact at any body link and automatically selecting stiffness are listed as future work. The project marks code as coming soon.
News
Origin discloses $20 million for interior-construction robots
October 15 announcement • Catch-up • Company-reported deployments
Origin emerged from stealth with $20 million in total funding, co-led by Accel and Zacua Ventures. Its OG-1 platform applies compound, sands and primes walls, with interchangeable tools intended to support additional interior trades.
The company says the robot uses force control to maintain sanding pressure and can work at heights up to fourteen feet. Crews still handle site mapping and material replenishment. That division of work matters when judging how autonomous a construction workflow really is.
Origin reports deployments across several types of building sites and plans larger international pilots. Its target of 150 deployed robots by the end of 2027 is a future goal. Claims of three-times-human speed and professional finish quality come from the company; the announcement supplies no independent throughput study or intervention-rate data. The funding supports expanding a working platform, while the economics at fleet scale still need evidence.
Highlights
Spark-0 has public code and a downloadable checkpoint
Resources accompanying the October 15 paper • Catch-up
The public repository includes training, data-processing and deployment code, and its linked model repository contains an actual pretrained weight file. This gives researchers more to examine than a demonstration video. The project still labels SparkArena-Sim code as coming soon, so the available release should not be described as a complete, independently reproduced benchmark package.
Read this first
Watch Compliant GMT’s paired stiff and yielding demonstrations. They make the value of adjustable compliance easy to see: the same intended movement can produce very different contact behavior. Read them as demonstrations of control, with human guidance in the manipulation examples.
Read the paper