CARE · Cooperative Action–Response Belief Estimation
Learning Action-Conditioned Team-Response Beliefs for Decentralized Multi-Robot Manipulation
Whether an action is useful often depends on how teammates respond to it. CARE predicts that response before acting, from local history alone.
RoboFactory · six collaborative tasks
85.8%
+13.3 over LatentToM, the best collaboration-specific baseline
Physical dual-robot tasks · 40 trials
92.5%
+35.0 percentage points over π0.5
Nominal action overridden
6%
of control steps, recovering half of the available headroom
Overview
Consider two robots carrying an object. One of them has to decide when to turn. Turning now shortens its own path, but it can destabilize the object if the partner is still regrasping. Waiting briefly makes little direct progress, yet it lets the partner finish the grasp so the team can proceed. The uncertainty that matters is what the team will do in response to the next action. The same decision arises when releasing an object in a handoff, yielding a shared workspace, or synchronizing a lift.
We call a robot's prediction of this action-dependent effect a team-response belief. For each candidate action, it describes the likely gain or loss in team utility that comes from letting teammates adapt, compared with teammates that keep their nominal behavior. It is conditioned on the candidate action, inferred from the robot's own history, and keeps its uncertainty. It needs no semantic label such as "the partner is ready" and no reconstruction of the partner's hidden intent.
Ordinary demonstrations provide almost no supervision for this belief. A successful trajectory shows the action that was taken and the response that followed, but nothing about how teammates would have responded to the alternatives. CARE (Cooperative Action–Response Belief Estimation) makes the response contribution an explicit training target through paired interaction augmentation, learns a predictive belief encoder that conditions nominal actions, then trains a candidate-conditioned response readout with the encoder and action generator frozen. Calibrated selection chooses an alternative or retains the nominal action at execution.
Same start, two choices
Method
A predictive belief encoder learns the interaction from local history and conditions the nominal action u0. The encoder and action generator are then frozen. K = 6 candidates include this nominal chunk and local behavioral alternatives such as waiting, retreating, and slowing down.
From a common state, each candidate is compared with the nominal action under two teammate behaviors: reactive teammates keep running their policies, replay teammates re-issue the commands recorded under the nominal action. The reactive contrast is the total advantage Ak, the replay contrast is the direct advantage Dk, and their difference Rk is the response label. The demonstrations stay in the training set.
A candidate-conditioned response readout uses the frozen predictive representation to predict quantiles of D, R and A plus an adverse-event probability for every candidate. Conformal calibration on held-out families turns the lower quantile of A into a lower bound. The best feasible alternative runs only if its bound clears the margin; otherwise the robot keeps u0. Execution uses local history only, with no rollouts.
One paired contrast, three labels
Utility U is a fixed weighted sum of progress, success, time to success and coordination penalties over a 16-step horizon. Collisions and drops are supervised separately as adverse events.
Identifiability
Two teammate behaviors can agree on every observed nominal continuation and still respond differently to an untried action, so nominal data alone cannot tell them apart. Matched reactive and replay branches turn Rk into draws from the belief, and the pinball loss recovers its conditional quantiles.
Decision value
Response information has decision value Γ(H) exactly when some candidate's relative response gain exceeds its direct disadvantage. A mean-based belief choice recovers at least Γ − 2e, where e sums the source-fidelity error and the estimation errors of the direct and response means.
Calibrated intervention
With exchangeable calibration and test families, the calibrated lower bound covers every candidate's realized advantage with probability at least 1 − α. A released alternative has paired utility above the margin δ on that coverage event. The guarantee concerns the paired reference distribution: source fidelity, exchangeability, and teammates departing from the reference policy at execution can affect deployment.
Results
Closed-loop success rate
RoboFactory (six tasks, two to four robots) and the MARS Control Track (four hidden-seed tasks). All policies share the same arm-local observations, action interface and control rate.
On RoboFactory, CARE achieves the highest average success among the evaluated policies. Most of the gain comes from Take Photo and Pass Shoe, where one robot must hold a pose or an object until its partner is in position and where maintaining support and waiting for a partner are central to completion. CARE falls behind on Place Food, Camera Alignment and Long Pipeline Delivery, where precise positioning or a long sequential chain dominates and there is less for the interaction contrast to change. On MARS, absolute performance is low for all policies; most of CARE's margin comes from Strike Cube Hard, whose success depends on a single well-timed joint impact, and CARE is the only policy with any successes on Four Robots Stack Cube.
Physical dual-robot evaluation
Successful trials out of 20 per task. Same local observations, action space, control frequency and task budget for every policy.
CARE, selector off
82.5%
the predictive encoder and nominal path remain active
Physical override rate
5.8%
selection adds four successful trials across both tasks
Folding Bed
20/20
the fold cannot continue until the partner has regrasped
Mechanism studies · fixed proposals
Ranking matters more than acting often
Fraction of oracle headroom recovered by each selector over the shared candidate set, with its override rate.
Explicit response supervision helps
Response mean absolute error, lower is better. The table reports the snapshot split and deployed seed. Total-only and Direct-only have no response predictor and are omitted from this error chart. All variants share proposals, representation, paired data, and selection rule.
Calibration turns interventions into gain
Closed-loop success against override rate for selection rules that share the same trained scorer.
Simulation · RoboFactory
Each arm receives only its own RGB history, proprioception, executed-action history and the shared task instruction. Teammate observations and simulator state are never available at execution. Frames are numbered in temporal order; the badge is CARE's closed-loop success over 20 rollouts.
MARS Control Track (hidden seeds, trained on its own 600 demonstrations): Place Cube in Cup 15% · Strike Cube Hard 95% · Three Robots Place Shoes 10% · Four Robots Stack Cube 20% · average 35.0%, the highest of all nine policies.
Real robots
Frozen policies evaluated on two physical two-robot tasks against ACT and π0.5. Both robots run the same CARE weights from their own observations. The advantage appears where the task cannot continue until the partner is ready, and shrinks where single-arm manipulation dominates.
Folding Bed · 20/20
The robots grasp the quilt edge, pull it over for the first fold, then regrasp for the second. One arm holds its edge and waits while its partner approaches and regrasps before the next fold. ACT succeeds in 12 of 20 trials and π0.5 in 9.
Opening Box · 17/20
The left arm opens the lid; the right arm takes over supporting it and waits while its partner grasps the target inside and places it on the table. π0.5 is comparatively strong here (14 of 20), and this is where CARE's few failures occur.
Opening Box
left opening the box
right waiting beside the box
left grasping the target
right supporting & waiting
left placing the target
right supporting & waiting
left moving back to home
right moving back to home
Folding Bed
blue grasping the quilt
green grasping the quilt
blue moving to fold
green moving to fold
blue waiting for green
green approaching & regrasping
blue moving to 2nd fold
green moving to 2nd fold
manipulating · supporting or waiting · moving
Scope and limitations
Team-response beliefs depend on what the local history reveals about the partner, on which actions are available and on how team utility is measured. Weak proposals leave little gain to recover, and a short horizon can miss delayed reactions. The paired contrast depends on the nominal trace and may put replayed teammates in states that never occur in ordinary interaction, so augmentation bias, unseen partners and policy-induced distribution shift can affect both belief accuracy and calibration. CARE already conditions nominal actions on predictive beliefs. Future work can adapt proposal distributions, extend response learning to changing teammate policies, and combine selection with runtime monitoring.
BibTeX
@misc{li2026beforeweact,
title = {Before We Act: Learning Action-Conditioned
Team-Response Beliefs for Decentralized
Multi-Robot Manipulation},
author = {Li, Zihao and Zhou, Yanan and Qiu, Ranpeng and
Ren, Guoqiang and Zhi, Weiming},
year = {2026},
note = {Preprint},
url = {https://jeong-zju.github.io/care/}
}