The search for a useful touch
A robotics preprint tests a training method for a quadrupedal manipulator that moves objects through contact rather than gripping them. It asks whether contact-guided exploration can overcome the exploration bottleneck and support robust task execution in non-prehensile locomanipulation. Here, that term refers to moving an object by pushing or otherwise making contact instead of taking hold of it.
The method uses candidate interaction regions from a general-purpose grasping algorithm, adds a dense exploration reward and gives exploration its own critic, or value estimate. At each episode reset, the target contact point is sampled from 25 candidates. The exploration signal is then faded during training: its weight falls from 0.1 to 0.01 between 5,000 and 10,000 training steps, while the regularization weight rises from 0.15 to 0.24 and the task weight stays at 0.75.
Simulation rewards contact, then task completion
Training took place in Isaac Lab across 4,096 parallel environments, with a 0.005-second simulation timestep and a 0.02-second control timestep. Each reset randomized the initial pose, joint states, mass and friction. For chair training, the assets combined 15 chairs from an IKEA furniture dataset with 100 procedurally randomized chairs.
Success in simulation meant getting the object within 0.2 metres of its goal. The evaluation counted a contact as missed when object displacement stayed below 0.2 metres, and defined tipover as more than 35 degrees. The paper says the reported trends were unchanged when the tipover threshold was varied from 30 to 40 degrees. Results were aggregated across five random seeds.
In the reported simulation comparison, the proposed method had a 94.1% success rate and a 4.4% tipover rate. Its completion time was 9.2 seconds, compared with 9.1 seconds for PPO + WS, another setup in the comparison. No confidence intervals or significance tests were reported for this comparison.
On chair transportation, standard PPO had a 9.1% missed-contact rate, while PPO + WS reduced that rate to 4.0%; the text describes learning under the single-value-function schedule as unstable and high variance. A version with no exploration reward had 0% success. Fixed-weight Multi-Critic PPO was described as solving contact but having a significantly higher tipover rate while exploration remained active.
The robot was less consistent on new objects
Hardware testing moved to the ALMA quadrupedal manipulator, using onboard sensing and external motion capture. Across four unseen IKEA objects, the policy succeeded in 40 of 58 trials, or 69.0%. Results differed by object: ADDE, 27 of 37 trials (72.90%); SANDSBERG, 8 of 14 (57.14%); VIHALS, 3 of 3 (100.00%); and LOVBACKEN, 2 of 4 (50.00%).
Object-level run counts were unequal, and no confidence intervals were reported. The 100.00% VIHALS rate came from three trials, while ADDE had 37.
Beyond the object sweep, the policy was reported to handle a moving goal, transport a chair with a total mass of 6.5 kg, and recover after an operator disturbance. Trial counts for those robustness tests were not reported. In a dishwasher task, the policy first used the handle to start opening, then switched to the door panel as the door's orientation changed, pushing it to a fully open horizontal position. That assessment was qualitative and did not include a success rate.
The paper also reports a 59% reduction in the fraction of timesteps during which the arm joint was within 10% of its position limit. The denominator and statistical uncertainty for that reduction were not reported.
A result still tied to the test setup
The numerical findings come with clear boundaries. Hardware validation used external motion capture, the dishwasher assessment had no quantitative success rate, and no confidence intervals or significance tests were reported for the simulation comparison.
The document is an arXiv preprint labeled arXiv:2608.28140v1 and dated 28 August 2026. It reports partial support from the Tuscany Health Ecosystem Next Generation EU project, Italian MUR/FISA and MUR projects, the Swiss National Science Foundation NCCR Automation, and the EU Horizon Europe programme.
Paper data and sources
Original title: Contact-Guided Exploration for Non-Prehensile Locomanipulation with Multi-Critic RL
Authors: Simone Tolomei, Mayank Mittal, Franco Angelini et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text