SafeBimanual: Safe Dual-Arm Trajectory Optimization

• SafeBimanual is a safety-guided trajectory optimization framework that leverages pre-trained diffusion models and gradient-based safety cost adjustments to refine dual-arm actions.
• It employs a vision-language model to dynamically schedule stage-dependent safety costs, addressing unsafe interactions such as gripper tearing, misalignment, and object collisions.
• Empirical evaluations show improvements up to 32.5% in success rate and 18.8% fewer unsafe interactions, all without retraining the underlying diffusion-based policy.

SafeBimanual is a test-time trajectory optimization framework for any pre-trained diffusion-based bimanual manipulation policies. It is designed for dual-arm settings in which two manipulators must coordinate over an action chunk $A_t=(a_{t+1},\dots,a_{t+n})$, with each action $a_i\in\mathbb{R}^d=(q_i^\ell,q_i^r,g_i^\ell,g_i^r)$ containing left/right arm joint positions and left/right gripper states. The framework addresses a central limitation of recent diffusion-based bimanual policies: they model multimodal action distributions well, but typically sample trajectories without explicitly enforcing physical safety, which can produce dangerous behaviors such as object tearing, gripper-object contact errors, and collisions between arms or manipulated objects (Deng et al., 25 Aug 2025).

1. Scope, formulation, and safety motivation

SafeBimanual is situated in dual-arm robotic manipulation for tasks such as dual-object pickup and transport, handover, stacking, pouring, carrying, wiping, and long-horizon household routines such as preparing breakfast. In this setting, the policy is conditioned on an observation $\mathcal{O}_t$ containing visual inputs, proprioception, and optionally task instructions, and predicts a finite-horizon action chunk rather than a single instantaneous command (Deng et al., 25 Aug 2025).

The framework is motivated by the observation that bimanual failures are often not caused by coarse task misunderstanding, but by unsafe coordination between two high-DoF manipulators. The paper identifies dangerous cases in which diffusion policies may generate trajectories where the two arms intersect or get too close, both grippers manipulate the same object inconsistently, object motions are spatially misaligned, grasps poke or scratch objects, or manipulated objects collide during cooperative transport. SafeBimanual therefore treats safety as a property of the sampled trajectory itself, not only of the learned policy prior (Deng et al., 25 Aug 2025).

A key design choice is that SafeBimanual is plug-and-play. It does not retrain the base diffusion policy. Instead, it uses the pre-trained denoising model as a trajectory prior and modifies reverse diffusion at deployment time through gradients of differentiable safety costs. This makes the method task-adaptive and stage-adaptive, which is important because the relevant safety constraints vary across tasks, object configurations, and phases of a manipulation episode (Deng et al., 25 Aug 2025).

2. Diffusion-guided trajectory optimization

The base policy follows a DDPM-style reverse process. At denoising step $k$, the reverse kernel is

$$p\bigl(A_t^{k-1} \mid A_t^k, O_t\bigr) = \mathcal{N}\!\left(A_t^{k-1};\, \mu(A_t^k, O_t, k),\, \Sigma\right),$$

where $\mu(A_t^k,O_t,k)$ is the denoising mean predicted by the diffusion model and $\Sigma$ is determined by the diffusion noise schedule. SafeBimanual preserves this structure, but augments it with a safety gradient computed from a scheduled cost $\mathcal{C}_{\mathrm{sched}}$ (Deng et al., 25 Aug 2025).

Because safety terms are difficult to evaluate on highly noisy intermediate samples, the framework first estimates a clean trajectory $A_{0\mid k}$ from the model’s noise predictor $\varepsilon_\theta$, and applies safety evaluation to that clean estimate. The guided reverse step is written as

$a_i\in\mathbb{R}^d=(q_i^\ell,q_i^r,g_i^\ell,g_i^r)$0

with $a_i\in\mathbb{R}^d=(q_i^\ell,q_i^r,g_i^\ell,g_i^r)$1 the denoising-step-dependent guidance strength and $a_i\in\mathbb{R}^d=(q_i^\ell,q_i^r,g_i^\ell,g_i^r)$2 the diffusion noise coefficient. Operationally, this is an energy-guided reverse diffusion procedure in action space: the pretrained model provides the nominal denoising direction, and the safety term pushes the sample away from unsafe bimanual trajectories (Deng et al., 25 Aug 2025).

The scheduled cost is a masked sum of candidate safety costs,

$a_i\in\mathbb{R}^d=(q_i^\ell,q_i^r,g_i^\ell,g_i^r)$3

where $a_i\in\mathbb{R}^d=(q_i^\ell,q_i^r,g_i^\ell,g_i^r)$4 is a binary scheduler output, $a_i\in\mathbb{R}^d=(q_i^\ell,q_i^r,g_i^\ell,g_i^r)$5 denotes relational keypoints, and $a_i\in\mathbb{R}^d=(q_i^\ell,q_i^r,g_i^\ell,g_i^r)$6 is the current task stage. The appendix describes the method as applying guidance in the final $a_i\in\mathbb{R}^d=(q_i^\ell,q_i^r,g_i^\ell,g_i^r)$7 denoising steps with $a_i\in\mathbb{R}^d=(q_i^\ell,q_i^r,g_i^\ell,g_i^r)$8, while the main text describes guidance in the final denoising steps with $a_i\in\mathbb{R}^d=(q_i^\ell,q_i^r,g_i^\ell,g_i^r)$9. In both formulations, the intent is the same: preserve sample diversity in early denoising and enforce safety only once the trajectory estimate is geometrically meaningful (Deng et al., 25 Aug 2025).

3. Unsafe interaction taxonomy and safety costs

A central contribution of SafeBimanual is a taxonomy derived from 1,320 demonstrations across 65 tasks and 7 benchmarks, in which over 96.2% of dual-arm unsafe interactions fall into five representative categories. The paper groups these into two broader classes: object unsafe interaction and gripper unsafe behavior (Deng et al., 25 Aug 2025).

Broad class	Unsafe interaction	Example
Object unsafe interaction	Object-object collision	Two manipulated objects collide while both arms are moving
Object unsafe interaction	Behavior misalignment	Bottle and cup are not properly aligned during pouring, insertion, or stacking
Gripper unsafe behavior	Gripper poking	A gripper approaches an object from an unsafe direction and scratches or pokes it
Gripper unsafe behavior	Gripper tearing	Both arms hold the same object but move inconsistently, causing unsafe stretching or shear
Gripper unsafe behavior	Gripper-gripper collision	The two end effectors approach too closely during handover or coordinated motion

The safety costs are defined on transformed task-relevant keypoints and gripper-tip locations computed from differentiable forward kinematics. The paper gives

$\mathcal{O}_t$0

where $\mathcal{O}_t$1 is differentiable forward kinematics, $\mathcal{O}_t$2 is the static offset from a grasped keypoint to the end effector, and $\mathcal{O}_t$3 is the fixed transform from end effector to gripper tip. Within one action chunk, the method assumes rigidity between end effector and grasped keypoint (Deng et al., 25 Aug 2025).

For object-object collision, the paper defines

$\mathcal{O}_t$4

This cost is minimized when the selected left and right object keypoints remain far apart, and is intended for patterns such as Dual Bottles Pick, Dual Shoes Place, and Place Fruits (Deng et al., 25 Aug 2025).

For behavior misalignment, the paper defines a structured relative-pose alignment cost. In the bottle-cup pouring example,

$\mathcal{O}_t$5

This penalizes misalignment orthogonal to a desired axis $\mathcal{O}_t$6 and deviation from a desired displacement $\mathcal{O}_t$7, thereby encouraging safe relational positioning between manipulated objects (Deng et al., 25 Aug 2025).

The framework defines five costs $\mathcal{O}_t$8, each tied to one unsafe pattern. The explicitly specified costs in the available formulation are $\mathcal{O}_t$9 and $k$0; the remaining costs correspond to gripper poking, gripper tearing, and gripper-gripper collision, matching the five-category taxonomy (Deng et al., 25 Aug 2025).

4. Vision-language scheduling and stage-dependent safety

SafeBimanual does not activate all safety costs uniformly. Instead, it uses a vision-LLM (VLM) to schedule which costs are relevant at each task stage. This scheduler is driven by the observation that the dominant safety hazard changes across a bimanual episode: a handover stage is dominated by gripper-gripper proximity, whereas a pouring stage is dominated by object alignment, and a jointly grasped-object transport stage may be dominated by tearing risk (Deng et al., 25 Aug 2025).

The appendix describes a two-part VLM mechanism. First, a CoT-VLM infers the most likely unsafe interaction pattern,

$k$1

from the current observation $k$2, relational keypoints $k$3, and task stage $k$4. Second, a scheduler VLM maps that inferred pattern to a binary mask,

$k$5

which determines which costs $k$6 enter $k$7 (Deng et al., 25 Aug 2025).

This stage-aware scheduling is central to the method’s practicality. A single fixed penalty cannot express the safety structure of a long-horizon bimanual routine. By contrast, SafeBimanual dynamically generates the relevant safety constraints for the current cooperation pattern by specifying keypoints and their pairwise relationships. The result is a form of test-time safety adaptation layered on top of a pretrained generative policy rather than baked permanently into model weights (Deng et al., 25 Aug 2025).

A plausible implication is that the framework is most effective in tasks whose safety hazards are relational and phase-dependent rather than purely local. This interpretation is consistent with the paper’s emphasis on dual-arm cooperation patterns such as avoiding tearing, avoiding arm-object collision, and maintaining object alignment.

5. Empirical results and practical behavior

SafeBimanual is evaluated on 8 simulated tasks in RoboTwin and on 4 real-world tasks. In simulation, it reports a 13.7% increase in success rate and an 18.8% reduction in unsafe interactions over state-of-the-art diffusion-based methods. In real-world experiments, it reports a 32.5% improvement in success rate. The paper presents these gains as evidence that safety-guided denoising improves both task completion and physical reliability rather than trading one against the other (Deng et al., 25 Aug 2025).

The method is also explicitly positioned as a remedy for common diffusion-policy failure modes in bimanual manipulation. These include trajectories where the two arms get too close, where both grippers manipulate the same rigidly grasped object inconsistently, where grippers poke or scratch objects, and where separate manipulated objects collide during coordinated motion. Because the guidance acts directly on predicted action chunks, SafeBimanual modifies the future trajectory before execution rather than only filtering already committed low-level commands (Deng et al., 25 Aug 2025).

Its deployment model is also operationally significant. Since the approach is test-time and policy-agnostic, it can be applied to any pretrained diffusion-based bimanual policy without retraining. This makes it suitable as a safety layer for systems that already use diffusion for multimodal action generation but lack explicit physical constraints.

6. Position within the broader safe bimanual manipulation literature

SafeBimanual belongs to a broader line of work that treats safety in dual-arm manipulation as a coordination problem, but it differs from earlier approaches in how and where safety enters the pipeline. Earlier kinesthetic-learning approaches such as SIMPLe encode safety through Cartesian impedance control, attractor and stiffness saturation, and uncertainty-aware stiffness regulation during interactive movement-primitive execution (Franzese et al., 2022). Collaborative control frameworks such as task-space SEIKO with fractal impedance control instead project operator intent into a feasible set defined by static equilibrium, joint limits, and contact constraints before execution (Wen et al., 2022). In teleoperation, other systems have enforced safety through whole-body velocity-damping inequalities and dynamically specified obstacle meshes generated from language-guided perception (Totsila et al., 7 Jul 2025).

Relative to those frameworks, SafeBimanual’s distinctive move is to make safety a trajectory-level test-time guidance problem for diffusion denoising rather than an impedance, quadratic-program, or teleoperation-filtering problem. Its costs act directly on sampled action chunks, which aligns it more closely with modern generative policy learning than with classical control synthesis (Deng et al., 25 Aug 2025).

At the same time, its limitations are structurally similar to those of other practical safe-bimanual systems. The framework improves safety likelihood and empirical robustness, but it is not presented as a formal safety certificate. This places it alongside contact-aware imitation-learning systems that improve robustness through phase estimation and compliant execution rather than invariant-set proofs (Ren et al., 2022). In that sense, SafeBimanual represents a specific synthesis: a generative policy prior, differentiable relational safety costs, and VLM-based stage scheduling, all assembled into a deployment-time optimizer for dual-arm coordination (Deng et al., 25 Aug 2025).