Robots Share Experience: A Fleet of 96 Machines Raised Packing Success from 79.2% to 98.3%

A team of authors has proposed a way for a fleet of working robots to train a single shared policy on its own, without people watching for every failure. On five real-world manipulation tasks, the resulting policies reached 93.3–99.2% success. Evaluation used 120 trials per task, on robots that did not take part in collecting the training data. In multi-stage phone packing, training with 96 robots raised success from 79.2% to 98.3%.

The underlying problem is familiar to anyone who has tried to move robots out of the lab. In operation, machines accumulate experience, and that experience is tempting to use to improve reliability. But the standard approach relies on people: a teleoperator watches the execution, steps in when something goes wrong, and those corrections go into training. With a dozen robots, that is tolerable. As the fleet grows, there are not enough operators to cover every machine. The authors call their solution collective experience learning, or CEL: deployed robots share their execution experience and improve a common policy.

How FERT works

CEL is implemented through a method called failure-guided experience retrieval fine-tuning, or FERT. The logic is as follows: when a robot fails, the system takes the context of the failure and searches a shared bank of successful experience from the whole fleet for trajectory fragments that fit the situation. The actions from the retrieved fragments become the training signal for updating the policy. No new expert demonstrations or human corrective actions are collected.

The second component, FATS, handles infrastructure. It is a fault-tolerant asynchronous training system that coordinates three processes: experience collection, policy optimization, and distribution of the updated policy to the robots. It is built on the assumption that part of the fleet will periodically drop out because of local failures. For real-world deployment, this is no small detail: machines stop, lose connectivity, or go in for maintenance, and training has to continue regardless.

What the experiments showed

Key figures from the paper:

• 5 real-world manipulation tasks, with final policies reaching 93.3–99.2% success;
• 120 trials per task, on robots excluded from training data collection;
• multi-stage phone packing: 79.2% → 98.3% when training with 96 robots;
• collection time per robot is held fixed, so fleet size directly determines the amount of experience;
• FERT generally delivers higher success across all fleet sizes tested.

Fixing the collection time per robot is especially important. It isolates the effect of fleet scale: each machine works for the same amount of time, but total experience is larger, and the policy benefits. The phrase "generally" in the original indicates that the improvement is not necessarily monotonic on every task and at every fleet size. Evaluating on robots outside the training set adds credibility: the policy did not simply memorize specific machines.

What to do with this in practice

For companies that already run fleets of manipulators in warehouses, assembly, or packing, the main takeaway is the economics of supervision. If improving a policy does not require constant work from corrective operators, the cost of raising reliability is decoupled from the number of people on shift. The fleet becomes a source of data, not just a consumer of it.

The limitations are also visible. The method relies on a shared bank of successful experience, so on tasks where robots almost never succeed, there will be nothing to retrieve. The evaluation covered five tasks, and transfer to other scenarios and robot types needs to be confirmed separately. The abstract gives no data on computational cost or network requirements, so it is worth reading the full paper before deploying anything. But the direction is clear: when there are many robots, it makes more sense for them to learn from each other than from a human looking over their shoulder.