MAWAM: A Multi-Agent World-Action Model for Coordinated Manipulation (Nov 20). The MAWAM preprint jointly predicts the next views and actions for a team of robot arms while preserving separate control interfaces. Coordinating the predictions lets the policy account for how one robot’s motion changes another robot’s next action.1
Readiness (Research prototype). The authors evaluated physical laundry-bag collection, towel folding, and umbrella handover, plus simulated tasks with up to four arms.
Pros. The authors report 67.8% mean success across the three physical tasks, up 18.9 percentage points over a baseline trained on the same demonstrations and cameras.
Cons. Performance on new workcells and object sets remains to be established.
Adapting Existing Policies
VG-Steer: Visual-Geometric Policy Steering from In-Context Aligned Abstractions (Nov 20). The VG-Steer preprint uses the geometry of one successful rollout to guide a frozen policy when an object moves outside its training positions. A manipulation team can test this as a way to extend a policy’s useful workspace without retraining it.2
Readiness (Research prototype). The authors evaluated pose and goal shifts in simulation and tested cable bagging on a physical robot.
Pros. They report success rising from 35.4% to 54.2% across cable placements on the physical test grid.
Cons. The physical trial used one tabletop setup, and success across other workcells is unknown.
Perception for Grasping
MIR-Depth: Multi-Range Refinement for Single-View RGB-D Depth Completion of Transparent Objects (Nov 20). The MIR-Depth preprint estimates transparent-object surface depth from one RGB-D view and an object mask, accounting for depth readings corrupted by refraction. Better surface estimates can improve candidate grasp poses for clear containers and similar objects.3
Readiness (Research prototype). The authors evaluated depth on five datasets and tested six-degree-of-freedom grasp pose prediction with a real robot.
Pros. The reported benchmarks show lower depth error than the compared methods.
Cons. The method needs an object mask, and the source describes public code and model release as planned.