Technical Summary: CoRMA (Contrastive Robotic Motor Adaptation)

Problem Statement

Contact-rich robotic assembly (e.g., PegInsert, NutThread, GearMesh) remains a significant challenge due to the sensitivity of these tasks to geometric mismatches, friction, compliance, and sensing noise. Unlike free-space manipulation, minor deviations can lead to jamming or irreversible failure. While recent sim-to-real systems like FORGE demonstrate robustness in simulation, they often degrade substantially when deployed on hardware. Existing approaches typically rely on task-specific training, demonstration-heavy pipelines (imitation learning), or classical adaptive control that requires repeated calibration for new setups. Furthermore, standard meta-learning frameworks like Rapid Motor Adaptation (RMA) are often designed for visual or object-centric cues, whereas contact-rich assembly requires interpreting force and proprioceptive histories to understand interaction states (e.g., free motion vs. jamming).

Methodology

CoRMA (Contrastive Robotic Motor Adaptation) proposes a context-based meta-adaptation framework tailored for force-dominant assembly. It modifies the standard RMA paradigm by replacing raw simulator-parameter adaptation with a compact, semantic contact context. The framework operates in three stages:

1. Semantic Contact Context Definition

Instead of using raw simulator parameters as privileged information, CoRMA defines a 6D Privileged $Z$ latent variable ($z_t$) that represents the semantic interpretation of contact dynamics. This vector includes:

• Contact onset
• Lateral engagement
• Guided contact transition
• Contact direction (along two axes)
• Jam or stick–slip tendency

These scores are computed from simulator contact and force dynamics and serve as supervised targets during training but are unavailable during deployment.

2. Training Pipeline

• Stage 1 (Privileged Teacher): A contact-aware policy is trained in simulation (Isaac Lab) using PPO. The observation space is augmented with the oracle Privileged $Z$, allowing the teacher to learn optimal behaviors conditioned on perfect semantic knowledge of the contact state.
• Stage 2 (Adapter Training): A causal Transformer adapter is trained to infer the Privileged $Z$ ($\hat{z}_t$) solely from deployable sensor histories (force, proprioception, and action). The adapter architecture includes:
  • Semantic Head: Performs regression to predict the 6D $Z$ vector ($L_{sem}$).
  • Contrastive Head: Uses an auxiliary InfoNCE objective ($L_{nce}$) to organize history embeddings based on coarse force regimes (free motion, first contact, guided sliding, jamming). This acts as a semantic regularizer, ensuring histories with similar interaction semantics cluster together, even across different tasks.
• Stage 3 (Deployment): The oracle $Z$ is removed from the policy. The frozen adapter predicts $\hat{z}_t$ online from sensor history, which is injected into the policy. This enables adaptation without demonstrations, privileged inputs, or test-time gradient updates.

3. Architecture

The core of CoRMA is a causal Transformer encoder that processes history windows of force, proprioception, and actions. It utilizes a learned readout token to generate the context representation. The use of a Transformer is highlighted as a dominant factor over standard Conv1D encoders used in previous RMA variants, allowing the model to capture long-horizon dependencies and sparse force transients.

Key Contributions

1. Semantic Contact Adaptation: CoRMA shifts the adaptation target from raw simulator parameters to a compact, reusable semantic contact vocabulary (6D $Z$) that describes interaction regimes rather than specific task geometries.
2. Force-Regime Contrastive Learning: The introduction of a contrastive objective based on force regimes (free, contact, sliding, jam) serves as a weak semantic regularizer. This helps structure the latent space to distinguish interaction types, improving the robustness of the context inference.
3. Non-Cheating Deployment Interface: The system adheres to the RMA principle where privileged information is used only for training supervision. At deployment, the robot adapts online using only onboard force and proprioceptive data, inferring the necessary contact context to handle real-world uncertainties.
4. Causal Transformer Adapter: The implementation demonstrates that a causal Transformer is superior to convolutional encoders for inferring contact semantics from long-horizon force histories.

Experimental Results

The authors evaluated CoRMA on three tasks (PegInsert, GearMesh, NutThread) in Isaac Lab/Isaac Sim 5.0 and on a real 7-DoF Marvin arm.

• Simulation vs. Real Gap: Strong baselines (FORGE) achieved near-perfect simulation success (98–100%) but suffered drastic performance drops in real-world deployment (e.g., FORGE dropped to 12.5% on PegInsert and 0% on NutThread).
• CoRMA Performance: While CoRMA's simulation success was slightly lower than FORGE in some cases, it retained significantly higher verified real success rates:
  • GearMesh: CoRMA 65.0% vs. FORGE 25.0%.
  • PegInsert: CoRMA 50.0% vs. FORGE 12.5%.
  • NutThread: CoRMA 59.3% vs. FORGE 0.0%.
• Adapter Analysis: Ablation studies confirmed that the causal Transformer provided the largest performance gain (increasing $R^2$ from ~0.43 to ~0.88 compared to Conv1D). The contrastive objective provided a smaller but consistent regularization benefit, improving representation quality without sacrificing numerical prediction accuracy.

Significance and Claims

The paper claims that semantic contact inference serves as a reusable adaptation interface within a related family of assembly tasks. By inferring a compact contact context from force history, CoRMA reduces the sim-to-real degradation observed in strong, task-specific baselines.

The authors are modest in their scope, stating that:

• The results support task-family-level meta-adaptation rather than unconstrained generalization to arbitrary unseen tasks.
• The current evaluation focuses on reuse within related tasks (PegInsert, GearMesh, NutThread) rather than broad generalization.
• Real2Sim calibration remains a future work item, as the remaining sim-to-real gap likely stems from mismatches in dynamics, friction, and sensor latency.
• The real-robot results are presented as controlled deployment evidence rather than full reliability certification, given the limited number of trials (tens per task).

In conclusion, CoRMA demonstrates that replacing raw parameter adaptation with semantic contact context inference, structured by contrastive learning, is a promising direction for robust sim-to-real assembly in force-dominant scenarios.