A new AI-driven robot system can assemble full-scale furniture using two arms and a model that tracks its own progress through the job.
Researchers introduced FurnitureVLA, a vision-language-action model fine-tuned on furniture assembly tasks broken into up to seven subtasks and as many as 1,550 control steps. Rather than treating assembly as one undifferentiated sequence, the system predicts a continuous progress signal alongside each action, so it knows where it is in the process and can transition between subtasks automatically. The team built a virtual-reality teleoperation rig to collect real-world demonstration data from a single human operator controlling both arms, then used a simulation pipeline to scale that data up. Tested on a real Kinova Gen3 dual-arm robot, the system hit 80% success in simulation across three furniture types, up from a 48% baseline, and held most of that performance on physical hardware.
Most prior robot assembly research either works at toy scale or relies on a single arm, which sidesteps the hard coordination problem that makes real furniture difficult. Getting two arms to cooperate on long, precise tasks without compounding small errors into catastrophic failures is exactly the kind of challenge that has kept warehouse-grade dexterity out of reach. The progress signal is the interesting design bet here — it gives the model a way to self-locate in a long task without being told explicitly.
A 16% performance drop moving from simulation to the real robot is notable: that gap tends to widen as tasks get harder, and furniture assembly in the wild involves surfaces, tolerances, and lighting that no simulation fully captures.