Generalist Robot Policies General Across Which Tasks, Robots and Environments?

A generalist robot policy maps observations and goals to actions across more than one task distribution. The claim should name the axes number of tasks, object diversity, environment changes, robot embodiments, sensors, language forms and task duration. A model can be broad on one axis and narrow on another.

A fixed library of scripted behaviors is not a learned generalist policy. A large language model that selects tools is a planner unless it generates grounded robot control. A policy trained on one arm can be a task generalist, but it should not be described as embodiment-general without transfer evidence.

Multi-task datasets combine trajectories with language or goal labels. Transformers model long multimodal sequences, while diffusion decoders represent several plausible action trajectories. Pre-training learns shared features and action patterns; fine-tuning aligns the policy with a target robot and task distribution. Prompting changes task intent but rarely resolves a new action space.

Few-shot adaptation uses a small target dataset. Zero-shot evaluation uses no target examples for the declared variable. Memory may be an observation window, recurrent state or external task record. Long-horizon policies often split planning from execution because one model call cannot reliably maintain state through many contacts and corrections.

Report a coverage vector tasks, objects, environments, embodiments, sensors and maximum uninterrupted horizon. For each test, identify what was held out and whether fine-tuning occurred. Use the same reset policy, intervention definition and number of trials when comparing baselines.

Measure task completion, unsafe contacts, recovery, latency and adaptation data. A policy that succeeds after a human resets every failed grasp is different from one that detects the drop and retries. Benchmark performance should be separated from deployment evidence such as repeated shifts or operation around workers.

RoboCat, RT-X, Octo and OpenVLA established multi-task and multi-robot learning paths. Octo and OpenVLA provide reusable open checkpoints, while RT-X emphasizes the shared data mixture. π0.5, GR00T N1.7 and Gemini Robotics 2 report broader household or humanoid capabilities with different access models.

Figure’s Helix is general across a growing set of tasks on one humanoid embodiment. That is technically meaningful but different from cross-robot generality. The distinction allows a strong task-general system to be described accurately without inflating its embodiment coverage.

Different robots expose different joint spaces, grippers, camera positions, velocity limits and control rates. End-effector deltas can provide a common task-space interface, but identical commands may produce different contact because reach, compliance and calibration differ. An embodiment token can tell the model which robot is active; it does not guarantee sufficient data for that body.

Cross-robot evaluation should include a held-out embodiment and state how much adaptation was allowed. Transferring semantic perception is easier than transferring precise motor behavior. A policy may recognize the correct object immediately while requiring substantial target demonstrations to grasp it.

Multi-task training can create interference improving one task reduces another. Rare tasks are underrepresented. Language aliases may map to different actions. Long tasks accumulate pose error and lose state after interruption. A policy may repeat a failed action because the observation window does not include the original goal or prior failure.

Recovery needs explicit training data, memory and safe exploratory actions. Failure datasets are often smaller than successful demonstrations because operators reset quickly. A generalist policy that cannot recognize uncertainty or request assistance remains fragile outside prepared test distributions.

Generalist policies are credible for flexible manipulation cells, research platforms and robots that need several related tasks with shared objects and tooling. They can reduce separate policy maintenance and improve adaptation to product variants.

Open-ended home assistance and unrestricted industrial autonomy remain experimental. Deployment should define an approved task envelope, monitor confidence and provide deterministic limits, remote support or safe stopping outside that envelope.