The dominant paradigm in robotic assembly has long relied on exteroceptive sensing—vision, force/torque sensors, or dedicated tactile skins—to guide contact-rich motions. This work questions whether a robot's own joint encoders provide sufficient signal for precise peg-in-hole tasks when environmental geometry is unknown. The core proposition is that high-resolution proprioception alone can support millimeter-scale insertion, gear meshing, and nut threading, provided the policy is trained on a sufficiently diverse suite of simulated assemblies.
Every modern robot carries joint position and velocity sensors as part of its control infrastructure. In legged locomotion and whole-body tracking, these signals have demonstrated sufficiency for dynamics regulation without visual feedback. The question this paper addresses is whether the same principle extends to assembly scenarios where contacts are finer and geometrically constrained. The authors argue that joint sensing, augmented by a short history of recent actions, carries enough information to close the loop for precise tasks, eliminating the need for visual or contact-based auxiliary sensors.
The policy architecture observes a short history of joint positions, velocities, and executed actions. No camera streams, force/torque readings, or tactile arrays are consumed. The part pose and fixture pose are both unknown at inference time, and the initial grasp configuration is randomized across a wide range. Under these conditions, the policy must discover insertion motions through physical interaction, using only the dynamics observable from its own joints. Reinforcement learning trains the policy entirely in simulation, where thousands of distinct assembly tasks are exposed, each with randomized part geometries, grasp poses, and contact configurations.
The simulation suite uses MuJoCo for physics simulation. The action space consists of joint-level torque commands. The reward function rewards progress toward the insertion goal while penalizing large slips or excessive forces that would indicate physical damage. Over the course of training, the policy discovers peg insertion, gear tooth meshing, and nut threading behaviors. The diversity of tasks—varying peg shape, hole diameter, gear tooth count, and nut pitch—ensures that the resulting policy cannot overfit to a single geometric configuration.
A key experimental result is zero-shot transfer to real hardware. The peg insertion policy, trained entirely in simulation, is deployed on a physical robotic arm with no further adaptation. Pegs of differing shapes and sizes are inserted successfully in 73% of trials. The authors attribute this transfer quality to the randomization regimen used during training: part poses, grasp orientations, and clearance dimensions are all randomized across a wide interval, forcing the policy to learn features that generalize across geometries rather than memorizing a specific assembly configuration.
In simulation, the authors compare blind policies (proprioception only) against matched visual policies that receive full state observation including part pose and depth imagery. The blind policies demonstrate notably better generalization to unseen part shapes and sizes. When presented with a part configuration never encountered during training, the blind policy succeeds at a higher rate than the visual policy, which appears to overfit to the visual features of training parts. The trade-off is that blind policies are typically slower, requiring more time steps to achieve insertion because they lack the immediate geometric cue that vision provides.
The data collection and task generation pipeline is a contributing artifact of this work. The authors release an open-source generator that procedurally builds assembly tasks for MuJoCo from exact convex primitives. Given a description of a fixture and a set of parts, the generator produces randomized valid assemblies with correct clearances. The authors report that this approach is both more accurate and faster than alternative methods such as convex decomposition or MuJoCo's signed-distance collision representation. The generator takes as input exact convex polyhedra and produces a valid simulated assembly in milliseconds, compared to the minutes required for offline decomposition pipelines.
The paper's findings suggest a reorientation of how sim-to-real transfer is approached for contact-rich manipulation. Rather than investing in additional sensors or carefully instrumented environments, the proprioceptive channel alone can suffice if the training regime emphasizes geometric diversity. The released task generator lowers the barrier for other researchers to simulate diverse assembly tasks, potentially accelerating research in blind manipulation policies.
Limitations noted in the results include the reduced speed of blind policies relative to visual baselines. The thousands of distinct assemblies required for training represent a non-trivial computational cost, though the procedural generator mitigates this for future work. The 73% success rate on real hardware, while promising, leaves room for improvement in reliability across broader part distributions.
In practice, the work implies that robotics engineers can reduce sensor payload weight and complexity by leveraging existing joint encoders for assembly tasks, provided the policy is trained with sufficient randomization. The open-source task generator offers an immediate tool for generating diverse simulated peg-in-hole, gear meshing, and threading tasks without manual scene setup.