LabDex Puts an XHand in a Real Chemistry Lab and Grades the Hierarchy
Lab robots usually get parallel-jaw grippers and a script. LabDex, from the University of Science and Technology of China with CUHK and iFLYTEK, puts a Franka Research 3 and an XHand on real glassware and asks whether a policy can climb from a single pour to a full experiment.
The paper (arXiv:2608.18618) and project page split the work into three levels: Atomic Skills, Compositional Skills, and Long-Horizon Laboratory Workflows. Real and simulated platforms share the same task names.
What is in the kit
The real cell uses three Intel RealSense D435i cameras (wrist, left-front, head), a Manus Quantum glove for the hand, and a VIVE Tracker for the arm. Demos are recorded at 20 Hz.
Atomic skills cover pick, place, insert, pour, press, shake, and stir on beakers, round-bottom flasks, graduated cylinders, test tubes, glass rods, funnels, and a weighing funnel. The paper instantiates more than 30 atomic skills on more than 20 objects, then groups them into more than 10 compositional skills and one long-horizon workflow: add NaCl from a weighing funnel into a round-bottom flask.
Default training uses 200 demonstrations per task. Evaluation is 50 trials each.
The scores
On atomic skills, π0.5 averages 0.57 success, ACT 0.34, Diffusion Policy 0.02. Placement is the easy bin (0.92 for π0.5). Pressing the heater is the hard one (0.12). The authors flag missing tactile sensing.
Compositional skills drop fast after the first grasp. π0.5 completes 0.73 atomic steps on average versus 0.52 for ACT. DP completes none. Putting a funnel onto a round-bottom flask is a typical cliff: the grasp works, the insert does not.
Nobody finishes the long-horizon NaCl workflow. π0.5’s average completed length is 0.62 atomic steps. The same insert after the funnel grasp is the bottleneck.
A 200 ml beaker pick that is 1.00 with no distractors falls to 0.00 when a similar 100 ml beaker is in view.
A Human’s Take
I am glad someone finally graded lab work as a stack instead of one “did the experiment succeed” bit. The message is blunt: foundation models can place a beaker and still cannot seat a funnel.
If you are selling an autonomous lab, start with the insert. The pour looks great in a GIF.