PACT boosts safe robotic GPU assembly with constraint gradients

stepjamUK · x · 2026-07-20

PACT tackles the problem of robots assembling GPUs and other delicate hardware by treating physical safety as a post-training alignment problem.

A pretrained diffusion policy is projected into a constraint-feasible region using only its own rollouts.
No demonstrations, reward function, or test-time guardrails are used.
The method distills a dense constraint gradient into every diffusion timestep and tightens constraints via curriculum learning.
Results: across 7 simulation bimanual benchmarks and 4 real-world tasks, safety violations fell by 31% on average while task success rose by 30.7%.
On the real GPU assembly task, safe rate improved by 36% and task progress by 47%, the largest gain among the tested tasks.

The main takeaway is that where the constraint gradient is injected matters more than simply adding more compute.

More from Embodied