The Missing Piece in Physical AI: Soft Actuation and Force-Aware Industrial Automation

In this talk

1. 0:02

   Why industrial robots need physical contact skills

   Kim says industrial automation still struggles to move from promising demonstrations to dependable deployments. Customers may begin with simple tasks, but they require reliability and safety. Robots must handle deliberate contact while controlling force, without sacrificing precision or payload.

2. 3:26

   Build both a compliant body and a force-aware policy

   Kim argues that safer contact requires changes to both the robot and its control system. Roboligent's Robin combines a soft actuation platform with a VLA that uses force information to adjust its behavior. The arm is described as reducing impact force by up to 10 times, with examples including sliding a glass plate, inserting a charger, pressing parts together, and lifting a box with both arms.

3. 6:05

   Pair focused AI with deterministic execution

   Kim says broad, general-purpose autonomy can have too low a success rate for industrial work, where tasks need predictable outcomes and coordination with equipment and sensors. Roboligent pairs a small VLA trained for skills such as gripping, picking, and screwing with execution logic such as finite-state machines or behavior trees. A demonstrated sensor task used 200 training episodes, or about 30 minutes of demonstration data.

4. 8:05

   Use force feedback to correct motion

   The policy combines a small vision-language model and a DiT action block with image, language, proprioception, and force data. A haptic flow-matching loss learns the expected force pattern for a successful task; deviations can prompt motion correction during execution. Kim says this approach cut interaction force by 70% versus a standard flow-matching model, and credits the work to Yuno Kim, a PhD student in Peter Stone's UT Austin lab.

5. 10:11

   Price around customer value and deployment

   In the audience discussion, Kim estimates Robin costs about $60,000 to $70,000 to build and could be offered to industrial customers for around $3,000 per month. He says U.S. robotics companies can compete by delivering enough ongoing productivity to justify the cost, with financing and staged deployment helping manage the upfront expense.