Artificial potential fields avoid obstacles in real time
At ICRA 1985 in St. Louis (25-28 March 1985) Oussama Khatib of Stanford showed that collision avoidance, until then treated as a high-level planning problem, can be distributed between levels of control: an obstacle raises an artificial potential field that repels the robot inside the control loop itself. Manipulator control is reformulated as direct control of motion in the space of the task rather than in joint space. The method was implemented in the COSMOS system on a PUMA 560 robot and demonstrated in real time on moving obstacles with visual sensing.
Why it matters
The robot gained a reactive layer: instead of recomputing a plan after every change in the world, as Shakey did, it answers an obstacle at the rate of the control loop. It is the same departure from sense-plan-act that Brooks would state a year later for mobile robots as the subsumption architecture, only done for an arm and with an equation rather than automata.
Two versions: the conference one (Proceedings of the 1985 IEEE International Conference on Robotics and Automation, pages 500-505) and the journal one (The International Journal of Robotics Research, volume 5, issue 1, Spring 1986, pages 90-98). The journal abstract adds that the field is made time-varying for moving obstacles, that outside the obstacles' regions of influence the end effector moves in a straight line with an upper speed limit, that the approach extends to all manipulator links, and that a separate joint-space field satisfies the internal joint constraints. The first page of the journal version mentions an earlier implementation on the MA23 manipulator in Montpellier in 1978. What the record does not claim: control rates, arm speeds, numbers of obstacles or any other figure; the journal copy on the author's site is a scan without a text layer, only its first page was read, and the conference full text is paywalled; the IJRR publisher's page sits behind a Cloudflare challenge.