How Many Demonstrations Do I Need to Train a Robot Manipulation Policy?
Training a robot manipulation policy requires a demonstration budget built around coverage. The right number follows from the task, sensing, architecture, deployment conditions, and starting model. These choices determine both how much the policy must learn and how much useful information each demonstration supplies.
Why "how many demonstrations do I need?" has no single answer
Every paper seems to give a different demonstration count, so practitioners have no steady basis for planning how much to collect. The disagreement is often one of scope. A result belongs to a particular task, architecture, and generalization target, none of which transfers automatically to another deployment. Task complexity and observation modality further change the requirement, as does the choice between fine-tuning and training from scratch.
How task complexity and structure set the floor
A reaching task with one natural action pattern gives the policy a different problem to learn from a task with several valid routes. Chi et al. (2023) show why this difference matters for diffusion-based policies, which must cover the modes of the action distribution. Going left or right around an obstacle introduces alternatives that a unimodal reaching task does not contain. The demonstrations need to show enough of both routes. This complexity raises the minimum count even before the robot meets new objects or works in new environments.
Contact richness raises that minimum for a different reason. Contact switches, precision requirements, and object variability expand the physical states the policy must cover. Wu et al. (Stanford, 2024) note that behavior cloning for contact-rich manipulation may require hundreds of demonstrations for a single task, while its scalability and generalizability remain unclear. Long-horizon tasks add to the burden because more primitives multiply the states the policy must cover. Their contact-retargeting approach makes single-demo transfer attractive precisely where collecting demonstrations for the full sequence becomes prohibitive.
Why demonstration diversity matters more than demonstration count
Once the task establishes a minimum, additional episodes that diversify the collection improve generalization. Lin et al. (ICLR 2025) found that generalization performance had a roughly power-law relationship with the number of training environments and objects, but not with the raw demonstration count. Beyond a per-environment or per-object threshold, more repetitions under the same conditions offered negligible gain. Coverage, rather than repetition, was the binding constraint.
The same study makes that distinction concrete. Data from 32 distinct environments, each with one unique manipulation object and a small, fixed number of demonstrations, led to roughly 90% success on new environments and objects for Pour Water and Mouse Arrangement. Using the same total demonstration budget in fewer environments did not lead to generalization. The mechanism is distribution coverage, because failures arise when the policy encounters states outside its training distribution. For practitioners, the first collection question is how many distinct contexts are needed. Even with a tight budget, the power-law shape favors collecting the first few distinct contexts because they improve generalization disproportionately more than contexts added later.
How observation modality changes the data requirement
Coverage also depends on what the sensor records. Richer geometric or physical observations reduce the relationships a policy must infer from examples. Ze et al. show that 3D Diffusion Policy (DP3), conditioned on sparse point clouds, achieves substantially higher average success across dexterous and table-top tasks than its 2D counterpart on the same demonstration budget. Depth shows geometric structure directly, so the policy has less to learn from RGB images. The episode count is unchanged, but the learning burden falls.
For contact-rich tasks, geometry alone does not describe the interaction. RGB video flattens force, contact, and compliance into pixels, leaving the policy to infer physical relationships indirectly. Tactile and force signals can instead record grip force and contact state that video cannot recover. EgoTactile (Tsinghua et al., ICML 2026) learns grasp pressure from egocentric video without instrumenting the hand at inference. ForceBand (Amazon FAR, 2026) combines egocentric video, wrist sEMG, IMU, and fingertip force in a ten-hour corpus for forceful manipulation.
These approaches change what an episode can teach. OSMO, an open-source tactile glove, records distributed contact for human-to-robot skill transfer and lowers the instrumentation barrier for force-enriched demonstrations. EgoDex (Apple, ICLR 2026) takes a different route by learning dexterous manipulation from large-scale egocentric video, bypassing robot-collected demonstrations during pre-training. Modality therefore affects more than sample efficiency. It can change where demonstrations come from and which physical states they describe.
How policy architecture sets its own data appetite
The model must turn that information into actions, and its representation sets another requirement on the budget. Diffusion Policy achieves strong single-task success with modest demonstration counts for well-scoped tasks, while ACT and similar transformer-based architectures operate in a similar range. Chi et al. identify coverage of multi-modal action distributions as a constraint on the required data. Architecture choice is therefore also a collection decision. The same task does not impose an identical learning burden on every model.
ConceptACT (Frontiers in Robotics and AI, 2026) provides a sharper example of this relationship. Episode-level semantic concept annotations describe what the demonstrator is doing, improving learning efficiency and performance with fewer demonstrations. The bottleneck can be missing structure in the representation of human intent. Demonstration quality interacts with that representation because expressive generative policies can reproduce pauses, dropped frames, and variable control frequency as features of the action distribution, increasing the effective data requirement.
How the generalization target multiplies the base count
A policy that succeeds with one object in one workspace has a narrower learning problem than one deployed across objects, environments, or embodiments. Expanding that target can require an order-of-magnitude more demonstrations for the same architecture and task. The deployment scope sets the multiplier on collection. RoboCasa and the Open X-Embodiment dataset illustrate the scale, spanning thousands to hundreds of thousands of trajectories as their targets extend across object categories, workspace configurations, and embodiments. Most single teams cannot collect that much data without a plan for doing so.
Internal data shows that teams that set their generalization target before budgeting consistently reach their performance thresholds with fewer total demonstrations than teams that first set a goal for the raw count. The practical rule from the scaling study is to budget demonstrations per distinct environment or object category first. Total count follows from those allocations. Observation modality then determines how well the demonstrations in each allocation capture the states needed for deployment, especially when contact information is essential.
How fine-tuning a foundation model changes the starting point
Fine-tuning changes what task-specific demonstrations must teach because a pre-trained model already contains generalized manipulation knowledge. New episodes primarily supply robot kinematics, camera configuration, and task specifics. Fine-tuning OpenVLA, Octo, or SmolVLA requires roughly 50–200 demonstrations per task. The OpenVLA technical report describes a 7-billion-parameter open-source VLA trained on Open X-Embodiment that supports parameter-efficient adaptation to new robot setups.
For a fixed, controlled deployment, a well-trained narrow policy with substantial task-specific data can outperform a fine-tuned foundation model. Generality can become a liability when the deployment distribution is tight. But the π0 technical report from Physical Intelligence shows how varied pre-training helps a model recover from mistakes in situations that are absent from the fine-tuning data.
How synthetic data and RL with demonstrations stretch a fixed collection budget
When people cannot collect enough demonstrations directly, synthesis can build on a small demonstration seed. MimicGen (NVIDIA/UT Austin, CoRL 2023) adapts human demonstrations to new contexts to produce large-scale, physically valid datasets. Its project documentation reports policies that match or exceed the generalization and success of policies trained on equivalent quantities of additional human demonstrations. The collection cost changes. The synthesized examples still need to supply the variation that deployment demands.
Hybrid imitation learning and reinforcement learning extend the seed differently. A 2025 study in Frontiers in Robotics and AI reports that demonstrations on the order of 20 significantly accelerate RL training and reduce total wall-clock training time. A 2025 paper identified as arXiv:2504.12609 claims that a single RGB-D human video can seed a sim-to-real RL policy for dexterous manipulation, using domain randomization over object properties within one task. Both routes depend on high-quality seeds because noise and inconsistency can propagate into the expanded data.
What demonstration count cannot fix: the contact-rich gap
The strongest scaling results do not resolve every manipulation problem. The scaling study establishes its laws on Pour Water and Mouse Arrangement while explicitly leaving more challenging dexterous tasks open. Those tasks may require substantially more demonstrations than its saturation thresholds suggest. Wu et al. likewise identify unresolved scalability and generalizability under distribution shift. Success in a fixed lab setting does not establish success when deployment conditions change.
Video-only or low-modality demonstrations can omit physical states that additional episodes of the same kind cannot supply. ConceptACT identifies another missing layer, namely the semantic knowledge humans use but low-level sensorimotor data alone does not represent. For surgical procedures, industrial assembly, and dexterous in-hand manipulation, force, compliance, and contact state are essential parts of the task.
A practical decision framework for budgeting demonstrations
To turn these constraints into a collection plan, answer five questions in order. First, define the generalization target before assigning an episode count. For a single fixed object and environment, use 100–800 demonstrations as the planning range for a narrow policy trained from scratch, or 50–200 for foundation-model fine-tuning. For cross-object or cross-environment deployment, budget for each distinct context and multiply by the number of contexts needed for coverage.
Second, check the task for contact-rich or dexterous requirements. If it is, set the sensing modalities needed before deciding how much to collect, since video-only demonstrations may fall short no matter how many there are. Third, decide whether to train from scratch or start with a pre-trained VLA. Fine-tuning lowers the task-specific floor by roughly four to ten, so the budget must account for this before plans are made for how to collect the data. Treat that reduction as a starting-point advantage while separately assessing the model’s suitability for a tightly controlled deployment.
Fourth, decide whether synthesis or RL can build on the seed collected by people. Simulation that matches the real task allows MimicGen-style expansion, while RL on the hardware can use a small set of demonstrations to speed up training. Fifth, set the standard for the quality of demonstrations before training. Require temporal consistency, low variance, and a fixed control frequency, with quality gates that keep noise out so the policy does not have to absorb it. The data collection specification template available on this site for preview and copy operationalizes these decisions through fields for generalization target, sensing modality, collection protocol, and quality criteria.