HiRe MoMa Adds Human Nudges To Boost TIAGo Robot Success On Messy Tasks

The short version

Mobile manipulators often fail when clutter or sparse training data breaks a single monolithic policy. HiRe MoMa keeps a frozen diffusion policy and learns tiny residual actions for the mobile base and arm, guided by sparse human corrections. On a TIAGo++ robot, success on a place task rose from 50 percent to 95 percent after about thirty minutes of online tuning.

What the researchers did

Zhefei Gong, Hefeifei Jiang, Zehao Wang, and Jiachen Li from Northeastern University, UC Riverside, and Georgia Tech posted preprint, not yet peer-reviewed HiRe-MoMa on arXiv on 14 October 2026, targeting CoRL 2026. The stack adds policy-agnostic residuals on top of a frozen Diffusion Policy. Residuals split into mobile-base and manipulator heads with one-way conditioning. A critic is reshaped using sparse human interventions covering less than twenty percent of training steps.

What they found

On TIAGo++ real hardware, Place Only success went from 50.0 percent to 95.0 percent versus the frozen base policy. Sweep Litter moved from 45.0 percent to 90.0 percent. Pick Place, the hardest task, rose from 22.5 percent to 40.0 percent. A flat residual baseline and +HG-DAgger variants lagged the hierarchical decoupled design. Average success across tasks moved from about 39 percent to 75 percent in project summaries, with execution about 1.29 times faster and human data under twenty percent coverage.

Why it matters for you

Warehouse and service robots need quick fixes on site, not full retraining. Residual heads let operators inject short human demos without throwing away an expensive base model. That is closer to how teams actually deploy on factory floors.

What to keep in mind

Tests focus on one robot platform and three household-style tasks. Preprint, not yet peer-reviewed. Pick Place success remains modest at 40 percent, so long-horizon grasping is not solved.

A mobile manipulator is a robot with wheels and an arm. It must drive and grasp in one job. Diffusion policies are a modern way to output smooth motion. Freezing the base policy means the team stops updating those core weights. Residuals are small add on corrections. Policy agnostic means the residual module can sit on top of different base policies.

Human interventions are short operator corrections during training. Sparse means they cover less than twenty percent of steps. The critic is a value model that scores states. Reshaping the critic nudges the robot toward states humans prefer. Hierarchical decoupling splits residuals for the base and arm. One way conditioning stops the arm head from fighting the base head.

Pick and place in clutter remains hard at forty percent success. That honesty matters for buyers. Teams should view HiRe MoMa as a fine tuning layer, not a full autonomy stack. Always pilot on your own floor layout before scaling fleet size.

Factories want robots that learn from a few human fixes. Residual learning adds small corrections. Base models stay stable. Human time stays limited. Reported gains are on one robot type. Your mileage may vary. Run pilots before fleet buys. Keep safety cages and e stops. Document who may intervene during learning.

Teams share papers so others can repeat the work. A preprint is an early public draft. Peer review may change claims later. Numbers in tables can move after feedback. Always read the linked paper for exact values. This desk note tracks when we last checked sources. We remove claims if sources disagree. If you build on this work, cite the primary paper. If you spot an error, tell the publisher through the site contact path.

Key takeaways

• Frozen diffusion policy plus learned residuals for base and arm.
• Sparse human interventions under 20% coverage reshape the critic.
• Place task hit 95% success after short online fine tuning on TIAGo++.
• Hierarchical decoupling beat flat residual and HG-DAgger baselines in tables.
• Posted 14 October 2026 on arXiv with CoRL 2026 target venue.

Photo: Pexels

Sources

• Primary — arXiv, Hierarchical Residual Policy Learning for Real-World Mobile Manipulation with Sparse Human Guidance (14 October 2026). TIAGo++ success tables and method diagram.