Tube Diffusion Policy for Contact Manipulation

• Tube Diffusion Policy is a reactive framework that combines diffusion-based imitation learning with tube-based feedback control to enhance action correction.
• It employs a two-phase approach where slow diffusion initializes actions and fast streaming refines them, addressing mismatches between control and high-frequency tactile updates.
• Leveraging a dual-time conditional U-Net for multimodal observation conditioning, TDP demonstrates robust performance in both simulated and real-world contact-rich manipulation tasks.

Searching arXiv for the cited TDP-related papers to ground the article in current arXiv records. Searching arXiv for the primary Tube Diffusion Policy paper. Tube Diffusion Policy (TDP) is a reactive visual-tactile policy learning framework for contact-rich manipulation that combines diffusion-based imitation learning with tube-based feedback control. In its canonical formulation, TDP uses diffusion to generate an initial action at chunk boundaries and a learned observation-conditioned feedback flow to update actions at high frequency within each chunk, thereby forming an action tube around nominal demonstration behavior. The method is designed to address a central limitation of chunked action-generation policies: they react only when a new chunk is generated, which is especially problematic when contact uncertainty, external disturbances, and high-frequency tactile sensing demand rapid closed-loop adaptation (Xue et al., 26 Apr 2026).

1. Scope and defining idea

TDP addresses visuomotor imitation learning in settings where observations combine proprioception with high-dimensional exteroception, particularly RGB vision and tactile sensing. The underlying motivation is that contact-rich manipulation requires both expressive modeling of nonlinear dynamics and rapid feedback. Diffusion-based policies provide the former, but most existing approaches rely on action chunking and therefore execute a sequence of future actions without intra-chunk feedback. In contact-rich settings, this creates a mismatch between policy update rate and sensory update rate, especially when tactile sensors operate at 100–150 Hz (Xue et al., 26 Apr 2026).

The defining move in TDP is to replace a purely open-loop chunk with an “action tube”: a neighborhood around a nominal action trajectory within which a learned feedback flow keeps the executed action close to the demonstration manifold while remaining observation-conditioned. Diffusion operates on a slow timescale to initialize the action at the start of a chunk, while streaming flow operates on a fast timescale to provide reactive, step-wise correction. The result is a hybrid controller in which diffusion supplies global correction and the tube supplies local stabilization.

A common misconception is that TDP is merely a faster variant of action chunking. In the formulation of the 2026 paper, that characterization is incomplete. The key novelty is not only reduced denoising cost, but also the replacement of open-loop chunk execution by a continuously updated, observation-conditioned flow in action space. The tube is therefore functional rather than purely geometric: it is realized by the behavior of the learned velocity field, not by a manually specified set.

2. Contact-rich manipulation setting

The method is formulated for continuous-control manipulation with observations

$$\mathbf{o}_t = (\mathbf{o}_t^{o}, \mathbf{o}_t^{r}) \in \mathbb{O} \subseteq \mathbb{R}^{n_o},$$

where $\mathbf{o}_t^{r}$ is robot proprioception and $\mathbf{o}_t^{o}$ is object-related information, which may be low-dimensional state or high-dimensional visual and tactile data. Actions are continuous control commands

$$\mathbf{u}_t \in \mathbb{U} \subseteq \mathbb{R}^{n_u},$$

and the system evolves under continuous-time dynamics

$$\dot{\mathbf{o}}(t) = f\big(\mathbf{o}(t), \mathbf{u}(t)\big).$$

The contact dynamics are unknown and highly nonlinear, with changing contact modes and frictional effects (Xue et al., 26 Apr 2026).

This setting makes action chunking particularly brittle. A chunked policy predicts

$$(\mathbf{u}_t, \mathbf{u}_{t+1}, \dots, \mathbf{u}_{t+H_a-1})$$

from the current observation and then executes that chunk without intra-chunk feedback. In manipulation involving slip, jamming, fingertip deflection, or local contact transients, small deviations accumulate rapidly inside the chunk. TDP is explicitly motivated by the need to preserve the expressive trajectory modeling of diffusion policies while restoring high-frequency reactivity.

The method also positions itself against two other baselines. Relative to streaming or flow-based policies, it adds a diffusion-based global correction mechanism; relative to classical tube MPC, it avoids explicit dynamics models and online nonlinear optimization. The paper therefore frames TDP as a learned analogue of tube MPC for visual-tactile manipulation.

3. Tube-based feedback formulation

The control-theoretic core of TDP begins from a nominal trajectory $(\bar{\mathbf{o}}(t), \bar{\mathbf{u}}(t))$ and a feedback law

$$\mathbf{u}(t) = \bar{\mathbf{u}}(t) + K\big(\bar{\mathbf{o}}(t) - \mathbf{o}(t)\big).$$

Differentiating and locally linearizing the dynamics around the nominal trajectory yields an action-space velocity field of the form

$$\dot{\mathbf{u}}(t) \approx \dot{\bar{\mathbf{u}}}(t) - K\Gamma(t)\big(\mathbf{u}(t) - \bar{\mathbf{u}}(t)\big).$$

Under the approximation $K\Gamma(t) \approx \lambda I$, this reduces to

$\mathbf{o}_t^{r}$0

This is the contracting tube dynamics: the action is driven forward along the nominal trajectory while being pulled back toward it (Xue et al., 26 Apr 2026).

In the learned version, a demonstration segment provides an interpolated nominal action trajectory $\mathbf{o}_t^{r}$1 over normalized trajectory time $\mathbf{o}_t^{r}$2. The target streaming velocity is

$\mathbf{o}_t^{r}$3

This decomposes into a feedforward term, which advances along the nominal trajectory, and a contraction term, which keeps the action inside the tube. The paper further defines a Gaussian neighborhood around $\mathbf{o}_t^{r}$4,

$\mathbf{o}_t^{r}$5

so the tube can be understood as a sequence of local action neighborhoods with time-varying width.

Theoretical analysis is given in discrete time. Under Lipschitz dynamics, bounded disturbance, bounded streaming imitation error, and a contraction assumption for the periodic diffusion correction, the tracking error is uniformly ultimately bounded. In the ideal case $\mathbf{o}_t^{r}$6, $\mathbf{o}_t^{r}$7, and $\mathbf{o}_t^{r}$8, the analysis yields asymptotic stability. This suggests that the method’s robustness claim is not only empirical but also tied to an explicit stability argument.

4. Network architecture and multimodal conditioning

TDP uses a dual-time conditional 1D U-Net. The network receives an action input and two distinct scalar time variables: diffusion time $\mathbf{o}_t^{r}$9 and trajectory time $\mathbf{o}_t^{o}$0. Separate sinusoidal embeddings are used for these two timescales, reflecting their different semantics. When $\mathbf{o}_t^{o}$1 and $\mathbf{o}_t^{o}$2, the network operates in denoising mode; when $\mathbf{o}_t^{o}$3 and $\mathbf{o}_t^{o}$4, it operates in streaming mode (Xue et al., 26 Apr 2026).

Observation conditioning is implemented through FiLM. A combined observation feature vector modulates residual blocks in the U-Net. The observation encoder stack depends on the sensing modality. For sim and real tasks, vision uses two RGB cameras, top-down and wrist-mounted, each encoded by a ResNet-18. Simulation tactile input is a 2304-dimensional vector from a dense tactile array and is encoded by an MLP. In the real world, Digit360 fingertip tactile sensors provide images, and each fingertip stream is encoded by a separate ResNet-18 with no weight sharing. Proprioceptive features are concatenated with vision and tactile features before FiLM parameter generation.

The tube is not represented explicitly as a geometric set inside the network. Instead, tube behavior emerges from the learned velocity field

$\mathbf{o}_t^{o}$5

which is trained to approximate the contracting action-space flow around demonstrations. This distinction matters: the method does not encode a hand-crafted invariant set, but learns a feedback field whose induced trajectories define the tube operationally.

5. Training objectives and online execution

Training is based on demonstration segments $\mathbf{o}_t^{o}$6 with observation-history length $\mathbf{o}_t^{o}$7, prediction horizon $\mathbf{o}_t^{o}$8, and execution horizon $\mathbf{o}_t^{o}$9. The total loss is

$$\mathbf{u}_t \in \mathbb{U} \subseteq \mathbb{R}^{n_u},$$0

For diffusion, the clean action is the action at the end of the observation history, and standard DDPM noising is applied:

$$\mathbf{u}_t \in \mathbb{U} \subseteq \mathbb{R}^{n_u},$$1

The denoising loss is

$$\mathbf{u}_t \in \mathbb{U} \subseteq \mathbb{R}^{n_u},$$2

For streaming, a trajectory time $$\mathbf{u}_t \in \mathbb{U} \subseteq \mathbb{R}^{n_u},$$3 is sampled, an action is sampled around the demonstration tube, and the network is trained to match the target velocity field:

$$\mathbf{u}_t \in \mathbb{U} \subseteq \mathbb{R}^{n_u},$$4

No auxiliary reconstruction or smoothness losses are introduced; the contracting behavior is induced by the target field itself (Xue et al., 26 Apr 2026).

Online execution alternates between two phases. At the start of each receding-horizon cycle, diffusion denoises a Gaussian action sample for a small number of DDIM steps to obtain an initial action. Then streaming updates that action at every small $$\mathbf{u}_t \in \mathbb{U} \subseteq \mathbb{R}^{n_u},$$5 using current observations:

$$\mathbf{u}_t \in \mathbb{U} \subseteq \mathbb{R}^{n_u},$$6

This division of labor is central. Diffusion need not fully solve local reactive control; it only needs to provide a coarse chunk initialization, after which the tube-based streaming phase supplies fast corrective feedback.

6. Experimental evaluation and efficiency

The empirical evaluation covers the Push-T benchmark, three virtual visual-tactile dexterous tasks, and two real-world experiments. On state-based Push-T, TDP achieves 96.1/96.9% average/max score with 106.2/105.1 steps, compared with 93.2/93.8% and 125.9/128.5 for DP (DDPM), 91.8/93.9% and 124.6/117.1 for DP (DDIM), and 88.6/89.7% and 124.4/116.1 for SFP. On image-based Push-T, TDP achieves 82.0/85.6% with 103.5/91.3 steps, again outperforming the reported baselines (Xue et al., 26 Apr 2026).

In the virtual dexterous tasks, TDP reaches 96% grasp success and 88% stable grasp success in stable grasping, versus 88% and 60% for DP (DDIM). In on-table reorientation, TDP reaches 90% success, compared with 82% for DP (DDIM), 80% for SFP, and 76% for TDP(stream only). In dish cleaning, DP with DDIM-10 attains 98.4% score at 72.9 ms latency, DP with DDIM-2 drops to 5.3%, while TDP with DDIM-2 attains 98.0% score with 13.3 ms denoising latency and 6.5 ms streaming latency.

Real-world results show the same pattern. On physical on-table reorientation, DP attains 60% success while TDP attains 96%. On jar opening, DP reaches 84% and TDP 96%. The reported latency for DP is 0.037 s per control, or about 27 Hz, whereas TDP reports 0.008 s for denoising and 0.003 s for streaming, yielding combined control above 100 Hz. Qualitative disturbance tests further show that TDP reacts to thumb bending during reorientation and to jar displacement during opening, whereas DP continues executing pre-chunked actions and often fails.

These results support a specific interpretation of TDP’s efficiency claim. The reduction in denoising steps is not presented as an isolated acceleration trick. Rather, the step-wise correction mechanism of the action tube allows diffusion to run with only 2–3 DDIM steps without major performance loss, because errors that would otherwise require more denoising are handled by the streaming phase.

7. Related formulations, terminological ambiguity, and limitations

The acronym TDP is overloaded in recent arXiv literature. In "VLM-TDP: VLM-guided Trajectory-conditioned Diffusion Policy for Robust Long-Horizon Manipulation," TDP denotes Trajectory-conditioned Diffusion Policy, where a voxelized 3D trajectory serves as a tube-like conditioning signal for long-horizon manipulation. That formulation treats the tube as an explicit spatiotemporal path in workspace coordinates rather than as an action-space feedback flow (Huang et al., 6 Jul 2025).

Two additional papers use the same acronym differently. "Efficient Diffusion Planning with Temporal Diffusion" defines TDP as Temporal Diffusion Planner, an offline-RL planning method that distributes denoising steps across the time dimension rather than a contact-rich policy with visual-tactile streaming feedback (Guo et al., 26 Nov 2025). "Targeted Data Protection for Diffusion Model by Matching Training Trajectory" uses TDP for Targeted Data Protection, and its “tube” interpretation applies to parameter trajectories during fine-tuning rather than robot control (Lee et al., 11 Dec 2025). These contemporaneous usages are conceptually adjacent through trajectory conditioning, but they refer to distinct objects: action tubes, workspace tubes, temporal planning tubes, and parameter-space tubes.

Within the specific 2026 Tube Diffusion Policy formulation, the reported limitations are concrete. The tactile representation is a simple concatenation of visual, tactile, and proprioceptive features and does not model sparsity, event structure, or asynchrony between touch and vision. Although denoising steps are reduced drastically, further acceleration is identified as a target for consistency models, one-step diffusion, or distillation. The tube formulation is described as generic and potentially applicable to flow-matching policies and VLA models with action chunking, but the experiments remain at moderate task horizons relative to full household manipulation (Xue et al., 26 Apr 2026).

Taken together, these points place Tube Diffusion Policy within a broader shift in diffusion-based robotics: from open-loop trajectory generation toward structured, closed-loop dynamical policies. Its distinctive contribution is to make the tube explicit in the control law itself—diffusion provides coarse global correction, while the learned observation-conditioned action-space flow realizes the tube as a reactive feedback mechanism suited to contact-rich manipulation.