Mixture of Frames Policy

Multi-Frame Action Denoising for Bimanual Mobile Manipulation

Dian Wang^*, Jisang Park^*, Xiaomeng Xu, Han Zhang, Shuran Song^†, Jeannette Bohg^†

^*Equal Contribution ^†Equal Advising

Bimanual Mobile Manipulation is Multi-Frame

A bimanual mobile robot naturally operates in multiple coordinate frames, including its base, its left and right end-effectors frames, etc.

The Choice of the Coordinate Frame Determines Action Representation Complexity

The same motion can be compact and pose-invariant in one frame yet spread into a broad, harder-to-learn distribution in another, so the chosen frame directly affects how hard the policy is to learn.

No Single Frame is Universally Optimal

Reaching is simplest in each hand's own frame, while holding a cup upright is better described in the base frame — so no single fixed frame is best for every phase of a task.

Mixture of Frames (MoF) Policy

Our Mixture of Frames (MoF) policy runs frame-specialized expert denoisers in parallel: it re-expresses one canonical diffusion state in each frame and fuses their noise predictions back in a shared canonical frame at every denoising step.

Simulation Experiments

We measure average task success rate (%) over nine tasks from BiGym (top) and DexMimicGen (bottom).

+15%p gap by frame choice

+16.5%p over Standard DP

+3%p over Oracle Frame Policy

F1.

No single hand-designed frame wins across tasks, and frame choice creates up to a 15%p gap.

Frame	Avg.	Best Tasks
Left	55.6	2
Right	55.0	1
Base	61.0	5
Rel Traj	55.8	1
Oracle / Worst	63.8 / 48.8	15.0%p gap

F2.

MoF outperforms both the oracle frame policy and the MoE baseline.

Method	Avg.
MoF-MoE	66.8
MoF-Ensemble	65.9
Oracle Frame	63.8
Single-Frame Ensemble	55.9
MoE-DP	55.1
DP	50.3

F3.

MoF's expert set, loss, and action representation are all critical.

Ablation	Avg.	Delta
MoF-MoE	66.5	0.0
Rel Traj canonical	64.0	-2.5
w/o Best Frame	60.3	-6.2
w/o Aux	58.5	-8.0
w/ Ortho Proj	59.5	-7.0

Q. Where does MoF-MoE gain over the Oracle Frame policy come from?

Router weight trajectories on the Threading task. — **Router weights across the Threading task.** Each curve is one expert's weight as the task progresses (faint lines: 20 individual episodes; bold lines: the 20-episode mean). The **rel-traj** (arm-centric) expert leads early, while each gripper reaches its own object — a motion that is simplest in an arm-relative frame. During the later coordinated bimanual insertion the **base** expert takes over, because aligning both arms requires consistent reasoning in a shared base frame. No single fixed frame fits both phases, which is exactly why MoF-MoE beats even the oracle frame here (+7.7%p).

Router modulation amplitude plotted against MoF-MoE advantage over oracle frame. — **The same effect across all nine tasks.** The horizontal axis measures how much the router varies its weights within an episode (within-episode weight std, averaged over experts); the vertical axis is MoF-MoE's success-rate advantage over the best single, or “oracle,” frame. More within-episode frame-switching predicts a larger gain — Threading modulates the most and gains the most, while tasks whose weights stay nearly constant see little benefit; Box Cleanup is the one task where MoF-MoE falls below the oracle frame. The improvement therefore comes specifically from *shifting frames as the task unfolds*, which a fixed-frame policy cannot do.

Real World Experiments

Our MoF-MoE outperforms the corresponding single-frame baselines.

Serving task rollout phases and success rates for single-frame baselines and MoF-MoE. — **Serving Task**

Pouring Task

MoF switches frames across phases.

MoF-MoE Router Weight Analysis

Our policy shifts its router weights across phases, prioritizing the left-frame expert during pouring because the static left frame provides a stable reference.

MoF-MoE (85%)