One policy, one pass/fail signal: all 20 NIST assembly tasks without per-task training
Synopsis
AlphaNIST trains a single reinforcement-learning policy across all 20 assembly tasks on the NIST Assembly Task Board 1 using only a binary success reward, with no demonstrations, progress shaping, scripted primitives, or per-task tuning. Prior learning-based methods typically solved one task at a time or relied on fixed sequences, staged initial poses, and dense reward guidance. In simulation, the shared policy reaches 98.2% overall training success and outperforms separately trained specialists on small-part tasks.
Interpretation
One policy can learn all 20 assembly tasks on the NIST Assembly Task Board 1, including nut threading, gear meshing, connector insertion, and peg insertion. Prior learning-based methods mostly assumed a fixed assembly sequence, constrained initial part poses to already-grasped or near-mating states, covered only a subset of the board, and often trained a separate policy per task. At 64K parallel environments over 25K iterations, the shared policy reaches 98.2% training success in 410 GPU-hours, while 20 separately trained policies average 81.4% and consume 16.1K GPU-hours in total.
With only a binary assembly-complete reward, the policy still acquires dexterous manipulation and planning-like behavior. Common practice adds dense progress rewards, demonstrations, or scripted subskills to make exploration tractable, whereas here intermediate progress receives no task reward. Across evaluations from 1K to 256K parallel environments, binary-reward success rises from 38.6% to 97.4%, versus 77.4% for dense reward shaping at the same scale; between 1K and 4K environments, dense shaping stays at 11.8% to 12.0%.
The same policy can complete up to ten assembly pairs in one episode without a prescribed order. Earlier long-horizon assembly usually relied on a high-level policy switching among imitation-learned options or a library of prescribed subskills, or completed pairs in a fixed order. Success is rewarded only when all active pairs are simultaneously complete, and success rates fall from 99% for one pair to 84% for ten pairs.
Perspective
All results come from simulation, with training and evaluation run in Isaac Lab's GPU-parallel environments, a 7-DoF Franka Emika Panda, and observations of robot state plus surface point clouds of the movable part, mating geometry, and robot. The policy outputs joint position targets and a gripper command, with velocity and torque limits set to match hardware. The authors plan to release code after review and revision, and state that transfer and repeatability will be evaluated on the physical board using sensor-derived observations and randomized contact and latency models.
The authors note that all results are in simulation, and that hardware-matched velocity and torque limits constrain the learned motions but do not establish robustness to perception error, latency, or contact-model mismatch. Extending the multi-pair framework to all 20 parts of the board remains an open goal, which the authors expect will likely benefit from more efficient simulation and learning methods under a single terminal task reward.