Diffusion policies teach humanoids to coordinate force and motion
Source published ·Modelwire updated
Original coverage: arXiv cs.LG ↗·How Modelwire adds context
The development
Researchers have developed FLORE, a diffusion-based policy framework that enables humanoids to learn complex manipulation tasks requiring coordinated force control across multiple limbs and the whole body. The system ingests vision, proprioception, and real-time force feedback to predict both motion and per-hand force commands, executed through a hybrid compliance controller that maintains balance during contact-rich tasks. This work bridges a critical gap in embodied AI: most prior learning approaches treat manipulation as pure kinematics, but real-world assembly, pushing, and grasping demand simultaneous motion and force regulation. The use of robot-free human demonstrations with wearable trackers and force-sensing grippers reduces data collection friction while scaling learning to full-body coordination, advancing the practical deployment of humanoids in unstructured environments.
Modelwire’s AI-generated summary of coverage from arXiv cs.LG.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The key innovation is not just learning manipulation from human demos (common now), but doing so while explicitly predicting and controlling per-hand force commands in real time. Most prior work treats contact as a binary constraint; FLORE makes it a continuous control variable that the policy must learn to regulate.
This work sits at the intersection of two recent threads in our coverage. The memory paper from October 22 showed that humanoids need temporal context to handle occlusion and dynamic scenes; FLORE adds the complementary insight that they also need force awareness to handle contact-rich tasks. Separately, the AdaDex framework (October 20) demonstrated that hybrid imitation plus goal-conditioned learning works for dexterous hands. FLORE extends that recipe to full-body coordination by treating force feedback as another modality the diffusion policy must predict, similar to how GPT-6-Astra (October 20) showed that tactile reasoning can be learned without specialized architectures. The practical payoff is clearer: robot-free human demos with wearable trackers reduce data collection friction, which aligns with the sim-to-real acceleration we saw in Isaac Lab 3.0 last week.
If the same FLORE architecture generalizes to contact tasks outside assembly (e.g., pushing through clutter, force-sensitive grasping on soft objects) without retraining the force prediction head, that confirms the approach captures generalizable force dynamics. If not, it's task-specific tuning masquerading as a general framework. Watch for a follow-up paper or benchmark release within six months that tests this on a diverse task suite.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.LG
AdaDex bridges imitation and reinforcement learning for robot hand skills
AdaDex addresses a fundamental tension in robot hand learning: reinforcement learning produces generalizable but unnatural behaviors, while motion imitation captures human-like dexterity but fails to adapt. This framework bridges both by training a shared policy on two data populations, bootstrapping from short mocap or video clips into flexible, multi-skill competencies. The approach matters because dexterous…
MentionsFLORE · diffusion policy · UMI grippers · humanoid robots
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “Learning Force-Reactive Humanoid Loco-Manipulation”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.