Bimanual manipulation policies trained with imitation learning are typically evaluated on workstation or datacenter-class GPUs, leaving the cost of deploying them on embedded hardware largely uncharacterized. We present a bimanual SO-101 system running entirely on an NVIDIA Jetson Orin Nano Super (8 GB), the entry-level tier of NVIDIA's embedded line, using a desktop GPU (RTX 3070) only for offline training, evaluated on pick-and-place of a deformable beanbag. First, we build a GStreamer capture pipeline backed by NVMM buffers that removes redundant host-device copies from three-camera sensing. Contrary to expectation, the conventional path fit the memory budget and dropped no frames; what zero-copy sensing recovers is CPU headroom (peak single-core utilization 98.0% to 77.0%) and worst-case latency (117.31 ms to 101.52 ms). Second, we train ACT and Diffusion Policy on identical demonstrations, each at its own reference budget (100k gradient steps for ACT, 200k for Diffusion Policy). ACT converges to a task-competent policy (19/20 trials) while Diffusion Policy does not converge to a usable one (0/10) even at twice the step count, which we attribute to differing convergence costs rather than an accuracy ceiling. Third, we convert ACT to TensorRT. FP16 reduces mean inference latency from 114.02 ms to 17.93 ms (6.4x) and INT8 to 12.65 ms (9.0x), with task success preserved at all three precisions (19/20, 18/20, 19/20). We report two findings not previously documented for ACT: TensorRT's general-purpose INT8 calibration quantizes the ResNet18 backbone but accepts zero of 145 transformer layers, explaining INT8's negligible size reduction over FP16 (0.9%) despite a further 28% latency gain; and the need for quantization is conditional on ACT's action-chunking configuration, feasible in full precision at n_action_steps = 100 but not at the per-step re-prediction temporal ensembling requires.
Real-world fine manipulation, particularly in bimanual manipulation, typically requires low-latency control and stable visual localization, while collecting large-scale data is costly and limited demonstrations may lead to localization drift. Existing approaches make different trade-offs: action-chunking policies such as ACT enable low-latency execution and data efficiency but rely on dense visual features without explicit spatial consistency, generative methods such as Diffusion Policy improve expressiveness but can incur iterative sampling latency, vision-language-action and voxel-based methods enhance generalization and geometric grounding but require higher computational cost and system complexity. We introduce a multistage spatial attention module that extracts stable 2D attention points and jointly predicts future attention sequences with a temporal alignment loss. Built upon ACT with a pretrained ResNet visual prior, a multistage attention module extracts task-relevant 2D attention points as a local spatial modality for action prediction. To maintain consistent object tracking, we introduce a self-supervised objective that aligns predicted attention sequences with visual features from future frames, suppressing drift without keypoint annotations and improving stability of the vision-to-action mapping under limited data. Experiments on simulated and real-world fine manipulation tasks, conducted on the ALOHA bimanual platform, evaluate task success, attention drift, inference latency, and robustness to visual disturbances. Results indicate improvements in localization stability and task performance while maintaining low-latency inference under the tested conditions.
Developing a unified policy for multi-task robotic manipulation remains challenging due to policy degradation from task interference and negative transfer. In this work, we propose Mixture-of-Experts-Enhanced Action Chunking Transformer (MoE-ACT), a parameter-efficient multi-task visuomotor framework tailored for bimanual manipulation. MoE-ACT incorporates sparse MoE layers into the ACT encoder, dynamically routing image tokens to selected experts based on task context, visual observations, and proprioceptive states. Furthermore, the framework incorporates task-conditioned Feature-wise Linear Modulation (FiLM) in the action decoder alongside multi-scale cross-attention, ensuring precise task grounding and capturing fine-grained spatial cues. Extensive evaluations on the RoboTwin 2.0 benchmark across 16 challenging bimanual tasks demonstrate that MoE-ACT achieves an average success rate of 62.0%, outperforming standard ACT by 17.4 percentage points. Crucially, with only 195M activated parameters, MoE-ACT surpasses the 16-fold larger foundation model π0 (3.24B) by 5.1 percentage points, exhibiting substantial gains in parameter efficiency and deployment feasibility. Real-robot dual-arm experiments consistently confirm its superior multi-task execution capabilities. Our open-source project page can be found at https://j3k7.github.io/MoE-ACT/.
Coordinated bimanual manipulation is challenging because the motion of either arm can alter the shared 3D scene and thereby affect the other arm. Yet most diffusion policies generate actions without explicitly modeling these future geometric consequences, while predictive variants typically use future state only as auxiliary supervision or fixed conditioning. We address this limitation by proposing JAMB, a diffusion policy that jointly denoises bimanual actions and future 3D point tracks. By allowing action and track hypotheses to evolve together within a shared Transformer, each can inform and refine the other throughout denoising. We further ground multimodal representations in a shared spatiotemporal coordinate system to facilitate geometry-aware interaction during joint denoising. We evaluate JAMB on diverse bimanual manipulation tasks in RoboTwin 2.0 and on a real-world robot, comparing it with action-only policies and alternative future-prediction approaches spanning different state representations and learning objectives. Across 16 simulation tasks, JAMB achieves an average success rate of 83.4%, outperforming the strongest baseline by 23.9 percentage points. On three real-world tasks, it outperforms the action-only and auxiliary geometry prediction methods by 50.0 and 21.2 percentage points, respectively. Beyond these performance gains, JAMB shows stronger generalization to cluttered scenes and out-of-distribution backgrounds than the evaluated baselines. Together, these results demonstrate the effectiveness of our joint action-motion modeling framework for coordinated bimanual manipulation. Our project website is available at https://jam-bimanual.github.io/