🤖 machine learning

CoWAM: Coordination Contracts for Selective Policy Intervention with WAMs

CoWAM introduces a selective intervention layer for bimanual robots that uses coordination contracts to safely modify World Action Model (WAM) policies only when specific synchronization, role, and collision criteria are met, significantly improving coordination success rates while maintaining a negligible risk of harmful interventions.

Original authors: Shuaijun Liu, Qifu Wen, Shuyang Hao, Qi Luo, Chenglong Zhang, Feiyang You, Chengyu Wu, Ningxin Su

Published 2026-08-04

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Shuaijun Liu, Qifu Wen, Shuyang Hao, Qi Luo, Chenglong Zhang, Feiyang You, Chengyu Wu, Ningxin Su

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a pair of robot arms to dance a complex tango. You don't just want them to move; you want them to move together. If one arm reaches for a cup while the other accidentally knocks it over, or if they both try to grab the same handle at the same time, the dance fails. This is the world of bimanual robotics, where two arms must coordinate perfectly to handle objects. To help robots plan ahead, scientists use something called World Action Models (WAMs). Think of a WAM as a "crystal ball" for robots: it takes a proposed move and simulates what the future looks like if that move happens. It's like a video game preview, showing the robot, "If you reach there, you'll knock this over." But here's the tricky part: just because the robot can see a future where things go wrong doesn't mean it should immediately change its mind. Sometimes the robot's original plan is actually fine, and panicking over a scary simulation causes more mistakes. The big question is: how do we tell the robot when to trust its gut and when to listen to its crystal ball?

This paper introduces a clever solution called CoWAM (Coordination Contracts with World Action Models). Instead of letting the robot's "crystal ball" randomly override its plans, CoWAM acts like a strict, fair referee that only intervenes when specific rules are broken. The researchers set up a system where the robot's original plan (the "nominal action") stays in charge unless a better alternative passes a series of tough tests. These tests are called coordination contracts. Imagine these contracts as a checklist of promises: "Promise you won't crash into the other arm," "Promise you'll hold the object at the right time," and "Promise you won't drop the cup."

The system works like this: The robot generates a list of possible moves, including its top choice and some alternatives. CoWAM looks at the "crystal ball" predictions for all of them. It only allows the robot to switch to a new move if that new move satisfies every single active contract (like no collisions, perfect timing) and offers a clear, low-risk improvement. If the new move is risky or doesn't clearly beat the original plan, CoWAM says, "Stick with the original plan." If even the original plan looks bad, the robot has a safety net to stop and ask for help.

The authors tested this in a simulated world with eight different two-handed tasks, like stacking bowls, lifting pots, and putting bottles in a dustbin. They found that CoWAM was much better at picking the right moves than other methods. Specifically, it successfully identified and chose valid, coordinated alternatives 93.3% of the time when an opportunity arose, compared to only 76.7% for a version without the full "contract" system. More importantly, it was very careful: it made harmful mistakes (like causing a crash) less than 1% of the time. In real-world simulations, this led to a 9.6 percentage point increase in successful task completion compared to the best previous methods. The paper shows that by using these strict "contracts" to filter predictions, robots can be both brave enough to try new things and smart enough to avoid disasters, turning a chaotic dance into a smooth performance.

Technical Summary: CoWAM – Coordination Contracts for Selective Policy Intervention with WAMs

Problem Statement

World Action Models (WAMs) enhance robot policies by pairing proposed actions with predicted future consequences. However, in bimanual manipulation, the mere visual plausibility of a predicted future does not justify overriding a nominal policy action (the top-ranked candidate, $i=0$ ). A coherent rollout may still fail due to subtle coordination issues such as delayed grasps, incompatible arm roles, or converging collisions. Conversely, overriding a nominal action without sufficient evidence can turn uncertainty into failure.

Current approaches often struggle to distinguish between a "productive-looking" candidate that hides a specific coordination violation and a genuinely superior alternative. The challenge is to develop a selective intervention layer that can rigorously evaluate whether an alternative candidate satisfies specific coordination obligations (synchronization, role compatibility, collision convergence) before replacing the nominal action, without retraining the underlying action generator or altering the candidate pool.

Methodology: CoWAM

CoWAM introduces a selective intervention layer that treats coordination requirements as coordination contracts. It operates on a frozen candidate pool provided by a WAM proposer and makes decisions to Preserve (execute the nominal action), Override (execute a verified alternative), or Abstain (invoke a fallback) based on a conjunctive set of conditions.

1. Coordination Contracts

The core mechanism defines obligations as typed contracts rather than a single plausibility score. Three contract families are identified:

Synchronization: Ensures temporal compatibility of contacts and releases.
Role Compatibility: Ensures arms retain task-consistent object and support roles.
Collision Convergence: Ensures inter-arm motion avoids unsafe contact.

A candidate $i$ is admissible only if it satisfies every active predicate $c_{i,e}$ for the current event $e$ . The contract-admissibility indicator $g_i$ is the product of these predicates. This structure ensures that a candidate cannot compensate for a specific coordination failure with a high utility score.

2. Event-Conditioned Verification

To resolve future-dependent failures that deterministic predicates cannot capture, CoWAM employs an event-conditioned verifier $f_\theta$ . Given an observation, candidate action, predicted future, and active event, the verifier outputs:

$\hat{p}_{i,e}$ : Estimated event satisfaction.
$\hat{r}_{i,e}$ : Estimated coordination risk.
$\hat{u}_i$ : Task utility.
$\hat{o}_i$ : Opportunity value.
$\hat{q}_i$ : Confidence.

These outputs are aggregated using an ensemble to construct conservative bounds for risk ( $R^+_i$ ) and utility ( $U^-_i$ ). A final score $S_i$ is computed combining utility, opportunity, and risk.

3. Calibrated Intervention Gates

An alternative candidate $i > 0$ is selected only if it passes a conjunctive set of gates ( $G_i = 1$ ):

Contract Validity: $g_i = 1$ (satisfies all active obligations).
Bounded Risk: $R^+_i \leq \tau_r$ .
Utility Retention: $U^-_i \geq U^-_0 - \epsilon_u$ (prevents discarding task progress for local safety).
Active Opportunity: $\hat{o}_i \geq \tau_o$ .
Confidence: $\hat{q}_i \geq \tau_q$ .
Selective Margin: $S_i - S_0 \geq \tau_m$ (must be decisively better than the nominal).

If no alternative passes, the system preserves the nominal action (if admissible) or abstains.

4. Outcome-Blind Same-Pool Evaluation

A critical methodological contribution is the evaluation protocol. All selectors operate on identical ordered candidate pools and commit their decisions before simulator outcomes are available. A shared oracle then labels the outcomes. This separates selector quality (the ability to choose the right candidate) from proposal quality (the quality of the candidate pool), preventing confounding factors in performance assessment.

Key Contributions

Coordination Contracts: A formalism specifying the typed obligations and evidence required for WAM-based intervention, ensuring that specific coordination violations are not masked by high utility scores.
Contract-Conditioned Selective Controller: A controller utilizing typed checks, event-specific verification, calibrated bounds, and a decision logic to preserve, override, or abstain.
Outcome-Blind Same-Pool Evaluation: A protocol that isolates selector performance from proposal generation, allowing for rigorous comparison of intervention strategies.
Robustness Evidence: Comprehensive analysis across component removals, candidate pool sizes, event families, sequential horizons, and different WAM proposer sources.

Experimental Results

The method was evaluated on eight bimanual tasks in the RoboTwin 2.0 simulator, covering cross-arm transfer, parallel placement, and multi-stage insertion.

Valid Selection: CoWAM achieved 93.3% valid selection (140/150 opportunities), a 16.7 percentage point improvement over the contract-only variant (76.7%) and significantly outperforming baselines like Future-Consensus (67.3%) and RGB-D selectors (84.0%).
Intervention Safety: CoWAM maintained extremely low error rates on matched negatives: 2.8% false interventions and 0.6% harmful interventions. In contrast, Future-Consensus had 62.2% false and 8.9% harmful interventions.
Closed-Loop Success: In natural closed-loop episodes, CoWAM achieved 62.9% success (151/240), a 9.6 percentage point gain over the strongest selective baseline (Selective Control at 53.3%).
Failure Reduction: CoWAM reduced inter-arm collisions by 66.7%, role conflicts by 75%, and asynchronous releases by 71.4% compared to the nominal Policy Top-1.
Robustness: Performance gains persisted across varying candidate pool sizes ( $K=4$ to $32$) and different WAM sources (X-WAM, LeWorldModel), demonstrating that the selector effectively utilizes diverse proposals without increasing error rates.

Significance and Claims

The paper claims that CoWAM establishes coordination contracts as an effective interface for conservative policy intervention. The significance lies in:

Separation of Concerns: By decoupling proposal generation from intervention, the method allows existing WAMs to be used safely without retraining.
Conservative Reliability: The conjunctive gating mechanism ensures that interventions only occur when a candidate satisfies all coordination obligations and provides clear, low-risk improvement, preventing "unsupported overrides."
Generalizability: The framework is shown to be reusable across diverse task structures and WAM candidate distributions, suggesting that coordination contracts can serve as a standard interface for world-action evidence in coordinated bimanual control.

The authors conclude that while prediction quality and intervention decisions are complementary, CoWAM successfully converts richer candidate pools into valid interventions without the error growth associated with unconditional reranking, thereby bridging the gap between predicted world-action evidence and reliable coordinated control.