When a two-armed robot is midway through a task and one target moves, today's learned policies usually regenerate the next stretch of motion for both arms at once. The brain appears to work differently. Primate motor cortex can prepare a new movement while the current one is still being executed, in separate, non-interfering directions of neural activity. Human hands correct for a jumping target by adding a new smooth sub-movement on top of the ongoing one, not by stopping and restarting. We propose a bimanual policy built on these two observations. Each arm keeps its own "preparation channel," a running estimate of the next correction that does not move the arm until a learned gate commits it. Committed corrections are added as smooth sub-movements to whatever the arm is already doing. Both arms read one shared map of where things are in space. The central empirical question is what should be shared between effectors and what should be kept separate: the spatial goal, the preparation, or neither.
This matters for robots and for neuroscience. For robots, it targets a specific failure: when one arm's plan changes, the policy regenerates the whole two-arm chunk, and the untouched arm can jerk or switch behavior. For neuroscience, the same model makes a testable prediction about an open question. Are the preparatory directions found in single-arm reaching private to each effector, or do they spill over across limbs? A cheap human bimanual target-jump experiment can check that prediction.
Core Claim
Mid-task corrections in bimanual manipulation are handled better by keeping preparation and triggering per arm while sharing only the spatial goal representation between arms. Corrections are then added as smooth sub-movements on top of ongoing motion, rather than regenerated as full two-arm action chunks. Compared with whole-chunk regeneration, this design should produce faster corrections, less unintended motion in the arm that was not perturbed, and no loss of success on tightly coupled two-hand tasks. A fully decoupled per-arm design should lose on those coupled tasks. A design that shares the preparation channel between arms should show measurable leakage into the other arm. The claim is falsifiable either way. If whole-chunk regeneration with latency masking matches us on the unperturbed arm, the architectural story is unnecessary.
Where This Sits in Prior Work
Three results motivate the mechanism. A Computational Mechanism to Account for Averaged Modified Hand Trajectories showed, in double-step reaching, that human hands correct by summing minimum-jerk motion units toward a shifting internal target. That model fit velocity profiles with normalized deviation 0.035, versus 0.078 for abort-and-replan. The inferred internal targets also matched saccade endpoints, which hints at a spatial goal shared across effectors. We reuse its superposition rule as the robot's correction primitive. Simultaneous motor preparation and execution in a last-moment reach correction task found that monkey motor cortex re-enters its preparatory subspace while movement activity continues. It also found a separate trigger subspace that governs commitment. That study was single-arm, so it leaves open exactly the question our per-arm versus shared-preparation comparison asks. When and why does motor preparation arise in recurrent neural network models of motor control? showed that preparation in "output-null" directions (activity that does not move the plant) emerges as the optimal solution for a single plant. We extend the idea to two plants that share one goal map.
On the representation side, Spatial Representations in the Parietal Cortex May Use Basis Functions showed that one population of gain-modulated units can serve several downstream effectors, reconstructing multiple reference frames with under 5% error. We use this as the design pattern for the shared goal layer: object locations are encoded once and modulated by each arm's own base and joint state. Specificity of Human Cortical Areas for Reaches and Saccades found graded rather than modular effector specificity in parietal cortex. That suggests the right answer lies between full sharing and full separation, which is why we treat the degree of sharing as the experimental variable.
Two results constrain the design. Single reach plans in dorsal premotor cortex during a two-target task showed that premotor cortex holds one plan per trial, not a blend: its dual-representation index was 3–6%, against roughly 85% predicted by a blending model. Our per-arm preparation channel therefore holds one candidate correction at a time, not a mixture. DECO: Decoupled Multimodal Diffusion Transformer for Bimanual Dexterous Manipulation with a Plugin Tactile Adapter found that naively concatenating input streams hurts bimanual policies, which warns against sharing everything.
The closest robotics work is Real-Time Execution of Action Chunking Flow Policies (real-time chunking). It treats inference delay as an inpainting problem: new chunks are generated so that they agree with the actions already committed. It runs smoothly on real bimanual tasks. That method still regenerates the joint action for both arms, with no per-effector preparation or commitment, and it was evaluated for latency rather than for target perturbations that affect only one arm. The work closest in spirit to the neuroscience, A Robotics-Based Approach to Modeling of Choice Reaching Experiments on Visual Attention, put dynamic neural fields (grids of units with local excitation and surround inhibition) on a two-joint single arm. It did not address multiple effectors.
Research Program
The policy has three parts. The first is the shared goal layer. A vision encoder produces object and target embeddings. These are multiplied by per-arm gain signals computed from that arm's proprioception, giving each arm its own view of one common spatial map.
The second part is a preparation network per arm. At every observation step, this small recurrent network updates a candidate correction: an endpoint displacement, a duration, and a coupling offset relative to the other arm. Its output does not reach the motors.
The third part is a per-arm trigger, a learned gate that commits the current candidate as a new motion unit. Each arm's commanded trajectory is a baseline chunk plus the sum of committed units:
$$x_a(t) = b_a(t) + \sum_k d_{a,k}\, s\!\left(\frac{t - t_{a,k}}{T_{a,k}}\right), \qquad s(\tau) = 10\tau^3 - 15\tau^4 + 6\tau^5 \text{ for } 0 \le \tau \le 1.$$
The symbols are defined as follows. $x_a(t)$ is arm $a$'s end-effector command. $b_a(t)$ is the baseline chunk from an ordinary action-chunking policy. $d_{a,k}$ is the displacement of the $k$-th committed unit, $t_{a,k}$ is its onset time, and $T_{a,k}$ is its duration. $s(\tau)$ is the minimum-jerk profile; $s$ equals 0 before onset and 1 after completion. Because committed units only add, a correction to arm $a$ cannot change arm $b$'s command unless the preparation networks are explicitly coupled. This is the property we want to test, not assume.
Training proceeds in two stages. First, an ACT-style baseline policy is trained on demonstrations. Then the preparation and trigger networks are trained on demonstrations augmented with synthetic target jumps. In these augmented demonstrations, the post-jump segment is relabeled by fitting minimum-jerk units to the residual between the original and the corrected path. Real demonstrations are not minimum-jerk, so the fitted units leave a residual. We report its size and keep a small learned residual head. If that residual carries most of the correction, the superposition story fails, and we will say so.
The scientific variable is the degree of sharing, compared across three variants. In the private-preparation variant, only the goal layer is shared. In the shared-preparation variant, both arms' candidates come from one network. In the fully decoupled variant, even the goal layer is separate per arm, in the spirit of DECO. We also measure the overlap between the two arms' preparation-state subspaces, using the preparatory versus movement variance analysis of Ames et al. This lets us report whether preparation in trained networks becomes effector-private on its own.
Baselines
The naive reference is synchronous action chunking with ACT from Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. The robot executes a chunk, pauses, and re-infers, which is effectively abort-and-replan. That paper reports 80–90% success on fine tasks such as battery slotting, but those figures are for unperturbed tasks. It gives a ceiling on clean performance, not on corrections.
The strongest competitor is real-time chunking with a flow policy Real-Time Execution of Action Chunking Flow Policies. It is the method most likely to already handle mid-motion changes smoothly. It reports motion 20% faster than synchronous inference and stable behavior under more than 300 ms of injected latency. No reported result was found on Lacuna for any method on bimanual target-jump perturbations, so we will produce all perturbation numbers ourselves. For fairness, real-time chunking gets the same baseline chunk network, the same demonstrations including the jump-augmented ones, the same inference latency, and the same hyperparameter-search budget measured in GPU hours. That way, any gap comes from the correction mechanism and not from extra data.
The decoupled competitor is a per-arm policy that follows DECO's separation principle. DECO reports 72.25% vision-only success on DECO-50, against 57.25% for ACT and 51.25% for Diffusion Policy. The ablations are the three sharing variants described above, plus a version in which triggered units replace the plan instead of adding to it. That last ablation isolates superposition itself.
Beating the strongest baseline means three things on perturbation trials. Correction latency must be lower, with no increase in unperturbed-arm deviation, at equal or higher success. On unperturbed trials, success must stay within the confidence interval of real-time chunking.
Why not just…?
Why not just use real-time chunking, which already corrects smoothly under delay? This is the real threat. Its inpainting keeps new chunks consistent with committed actions, and it may treat a one-arm target jump well enough. Two considerations cut the other way. It regenerates both arms jointly, so a jump seen by one arm can change the other arm's future actions. And its masks are built for latency, not for detecting a perturbation. The perturbation suite settles this directly. If real-time chunking shows unperturbed-arm deviation and correction latency comparable to ours, we conclude the per-arm preparation machinery is unnecessary and report that.
Why not just run two fully separate single-arm policies? Separation removes leakage by construction. DECO's gains suggest that decoupling inputs helps. But handovers, cable insertion, and holding-while-manipulating need each arm to know the other's goal. We predict the fully decoupled variant matches us on independent-arm tasks and loses on coupled ones. If it does not lose on coupled tasks, the shared goal layer is not earning its place.
Why not just re-plan at high frequency with a small residual correction network? Fast closed-loop re-inference with a learned residual is simpler than preparation plus triggering. Our add-instead-of-replace ablation, combined with an ungated residual head trained on the same augmented data, is exactly this alternative. If it matches the gated version, the trigger adds nothing. Ames et al.'s finding that commitment briefly locks the plan predicts some benefit from gating, but that is a prediction, not evidence for robots.
Proposed Experiments
The first experiment is a simulation perturbation suite, extending tasks from RLBench: The Robot Learning Benchmark & Learning Environment to two arms. Targets jump at onsets spanning early to late in each reach, mirroring the 10–110 ms interval sweep of Henis and Flash. Tasks split into independent-arm and coupled-arm groups. Metrics are success, correction latency (time from jump to first deviation toward the new target), unperturbed-arm path deviation, jerk, and mode-switch rate. We also run the single-effector Kinetix tasks from the real-time chunking paper as a sanity check that we have not lost latency robustness.
The second experiment is on real hardware. We use an ALOHA-style platform from Mobile ALOHA: Learning Bimanual Mobile Manipulation using Low-Cost Whole-Body Teleoperation and repeat a subset of real-time chunking's real tasks, such as cable insertion, with experimenter-induced object displacements.
The third experiment is offline. On DECO-50, which provides separate left and right action streams, we train the three sharing variants and measure held-out action prediction per arm. We also measure cross-arm predictive information, which tells us how much natural bimanual demonstrations favor shared goals. We do not claim to reproduce DECO's real-robot 72.25% without the Unitree hardware.
The decisive checks are three. The private-preparation variant must beat the shared-preparation variant on unperturbed-arm deviation. It must beat the decoupled variant on coupled tasks. It must match or beat real-time chunking on correction latency.
The fourth experiment is a small human bimanual double-step study (on the order of 16 participants, as in the grasp–articulation work). One hand's target jumps, and we track the other hand at high frame rates, as in Continuous action with a neurobiologically inspired computational approach reveals the dynamics of selection history. We analyze single trials, not averages, following the caution from the single-plan result. If the unperturbed hand shows graded, onset-dependent deviation, human preparation looks more like our shared-preparation variant than the private one. That finding would inform which robot design is closer to biology, and it would address the open question of effector specificity left by Ames et al.
References
- A Computational Mechanism to Account for Averaged Modified Hand Trajectories
- Simultaneous motor preparation and execution in a last-moment reach correction task
- When and why does motor preparation arise in recurrent neural network models of motor control?
- Spatial Representations in the Parietal Cortex May Use Basis Functions
- Specificity of Human Cortical Areas for Reaches and Saccades
- Single reach plans in dorsal premotor cortex during a two-target task
- DECO: Decoupled Multimodal Diffusion Transformer for Bimanual Dexterous Manipulation with a Plugin Tactile Adapter
- Real-Time Execution of Action Chunking Flow Policies
- A Robotics-Based Approach to Modeling of Choice Reaching Experiments on Visual Attention
- Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware
- RLBench: The Robot Learning Benchmark & Learning Environment.
- Mobile ALOHA: Learning Bimanual Mobile Manipulation using Low-Cost Whole-Body Teleoperation
- Continuous action with a neurobiologically inspired computational approach reveals the dynamics of selection history