Paxton says model-based control has no future
A question about what RL adds over inverse kinematics turned into a call to drop online optimization altogether.
Chris Paxton put a hard line on the running argument about reinforcement learning versus classical control. "The future clearly is not in model based control, even for wheeled robots," he posted. That is a broader claim than the one he started with, and it covers machines nobody usually frames as a balance problem.
The future clearly is not in model based control, even for wheeled robots
The thread began when Xiaolong Wang asked Paxton a simple question about a Dyna demo. What does RL add here compared to doing IK (inverse kinematics, solving for the joint angles that put the hand where you want it). Paxton's answer was reactivity. He pointed to many object masses, to a sudden impact, to an arm getting caught, and said whole body RL handles those more gracefully than IK does without a lot of engineering effort.
Liao goes further
Qiayuan Liao then widened it to everything. Any online optimization for low-level control, IK, ID (inverse dynamics) or MPC (model predictive control, replanning against a model every timestep), should be abandoned. His reasoning is that an RL policy can track more precisely, more robustly, and adapt to unknown disturbances and uncharacterized noise. He said that holds especially for low-cost hardware, and even on a manipulator, which is the case classical control people usually consider settled.
That is the part worth sitting with. Arms with decent encoders and known payloads are the home turf of model-based methods. Liao is saying the cheap-hardware argument applies there too, because the model you are optimizing against is only as good as your characterization of the robot, and on inexpensive parts it is not good.
What the demo actually shows
Paxton conceded the weak point himself. As others point out, he wrote, Dyna does not show any of these things, so it is less clear. He still thinks it is the right strategy. So this is a bet about where effort should go, not a measured result. Nobody in the thread posted a comparison where an RL controller beat an IK stack on impact recovery or on a caught arm.
This is the same fight that has been running all week in earlier rounds, and it keeps landing in the same place. The failure cases everyone argues about, odd payloads, collisions, hardware that does not match its model, are exactly the ones demos rarely put on camera. Until somebody films them, the strongest claims in this debate stay claims.
Any online optimization (IK, ID, MPC) for low-level control should be abandoned.
As others point out, Dyna doesnt show any of these things, so its less clear. But I do think its the right strategy
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.

