OpenAI's robots learn real-world tricks without real-world training

OpenAI has cracked a persistent robotics problem: getting simulated training to actually work on physical hardware. The secret isn’t making simulations more realistic — it’s making them deliberately chaotic.

The team developed a technique called dynamics randomization that deliberately messes with 95 different physical properties during training. Mass, friction, sensor noise, action latency, table height — all of it gets randomized. An LSTM-based policy learns to use past observations to figure out the actual dynamics it’s dealing with and adjust on the fly. Feed-forward networks couldn’t handle this at all. The LSTM can.

They also tackled the reward problem head-on. Most robotics tasks don’t have neat, incremental rewards. Stacking a block requires encoding arm proximity, orientation, grip, lift height, and placement distance — a nightmare to hand-craft. After months of failing to get conventional reinforcement learning algorithms to do pick-and-place, the team built Hindsight Experience Replay. HER lets an agent learn from binary success/failure signals by reframing failures as intentional outcomes. You aimed for a gas station, ended up at a pizza shop — you still don’t know where gas is, but now you know where to get pizza. Combined with visual domain randomization, the system maps vision directly to action without special sensors.

The computational cost is real. Dynamics randomization slows training by 3x. Learning from images rather than simulator states adds another 5-10x. But the result is a closed-loop system that reacts to unplanned changes — a genuine departure from the brittle open-loop systems that came before. OpenAI sees three paths to general-purpose robots: massive physical robot fleets, hyper-realistic simulators, or randomized simulators that force generalization. They’re betting hard on option three.