Researchers from Southwest Jiaotong University, the universities of Leeds, Oxford and Surrey, and Southeast University have reported a robot controller that evaluates possible movements before inserting a flexible rod. Across 90 fixed starting configurations in a physics simulation, RodForesight raised overall insertion success from 88.9% for the same two-stage diffusion controller without its predictive model to 96.7%. The September 10 preprint has not been peer reviewed.
Rigid peg-in-hole controllers infer the tip from the gripper because the peg keeps its shape. A flexible rod bends during contact, so its stiffness changes the tip angle and lateral position. The simulated task placed a rod 100 millimetres long and 4 millimetres wide above a 5-millimetre hole. The researchers split correction of the initial pose from the smaller movements near the hole.
RodForesight begins with side and wrist RGB-D camera views, which contain colour and depth information. A visual-servoing controller uses masks of the rod and hole to bring the tip into a small hand-off region. A diffusion policy then proposes eight sequences of 16 gripper movements. A gated recurrent unit, or GRU, world model predicts the radial and tilt errors after the first four movements in each sequence. The controller selects the proposal with the smallest predicted tilt, uses radial error to break a tie, executes four movements and takes another observation before planning again.
The team trained the diffusion policy on 300 expert insertion trajectories, producing 6,654 image-and-action samples. It then simulated 53,232 alternative action branches from matched starting states to train the world model. The principal test covered stainless steel, copper and carbon-epoxy rods, while every failed attempt remained in the denominator. A trial counted as successful only when the rod entered at least 15 millimetres while staying within 0.5 millimetres of the hole centre and 2 degrees of tilt. All images, contacts and material responses in these tests came from a Cosserat-rod simulator.
The two-stage diffusion controller completed 88.9% of the principal trials without a world model and 96.7% with the GRU predictor; a paired McNemar test gave a p-value of 0.0156. On 256 cases with unseen hole positions and a wider range of starting poses, the complete controller succeeded 87.5% of the time. Without retraining, it also reached 94.4% on simulated aluminium and dry PA66 rods whose stiffness lay below the training range.
Performance weakened when rod diameter changed. A controller trained on 2-, 4- and 6-millimetre rods succeeded in 68.9% of tests with a 3-millimetre rod and 86.7% with a 5-millimetre rod; perception recovery accounted for 89.3% of failures in the thinner-rod test. The world model's radial mean absolute error also rose from 0.028 millimetres on matched test branches to 0.576 millimetres on states reached during principal trials and 0.761 millimetres in the broader-pose test. Physical trials now need to show whether the same camera-to-action loop can handle real lighting, friction, material variation and contact without damaging the rod or hole.
Sources
- Chuanbo Yu and colleagues, arXiv, September 10, 2026. Abstract, authorship, submission date and headline result.
- Full paper. System sequence, simulation setup, controller comparisons, generalisation tests and limitations.
- Lead image: Larry D. Moore, “Tour4 Spoetzl Brewery Shiner Texas 2024”, under the CC BY 4.0 licence. The photograph is illustrative and does not show the experiment.
