Recap: 1,816 Parcels per Hour at 98%+ Success Rate | Robotic Logistics Sorting Live Challenge

In this talk

1. 0:42

   Designed for a working sorting line

   The robot handles boxes and soft parcels, turning items when needed so their barcodes face up. It also diverts irregular parcels, such as a foam cooler or a rolling cylinder, to a separate chute, following the line’s five-sided scanning process.

2. 3:59

   A fast, adaptable two-arm system

   The setup uses two arms, three cameras, and a VLA model that identifies parcels and plans how to handle them. The team chose grippers over dexterous hands to keep equipment affordable and easier to repair; the grippers can be swapped in under a minute. They said a station can be deployed or adapted to a new line in about 30 minutes.

3. 5:46

   Balancing speed, success, and recovery

   The team said production-line success rates were already around 95% to 98% or higher, with errors often involving wrinkled parcels or items overhanging a transfer belt. The robot sometimes recovers when a parcel slips or gets stuck, and the team uses difficult cases and reinforcement learning to improve performance. The stated longer-term target is 99.99% success.

4. 29:20

   Deployment and model development

   The team described scaling to hundreds or thousands of stations as a major challenge, especially keeping hardware and model performance consistent across different lines. Zibianliang is also exploring education, logistics, and home robotics, aiming to improve one foundation model through deployment feedback and post-training. The lead said inference takes tens of milliseconds.

5. 1:02:05

   Live challenge result

   The team finished the one-hour challenge with 1,816 parcels sorted at a 98% success rate. Professor Gan said the result was comparable to a skilled human worker and reflected the team’s model and data pipeline; the host said it represented normal operating performance, not a peak result.