Low-volume training for moving agents: Dynamic Intelligence vs Physical Intelligence, Figure AI, Skild, and Tesla Optimus

Across these organizations, the destination looks similar: policies that generalize across tasks, scenes, and embodiments rather than brittle single-skill controllers. Moving agents must perceive, plan under contact and dynamics, recover from failure, and transfer skills when the environment shifts.

Where strategies diverge is the training diet. Some programs optimize for massive multimodal corpora, fleet video, and broad teleop coverage. Others optimize for fewer, higher-signal trajectories, stronger evaluation gates, and RL loops that improve behavior with little or no new demonstration data. Dynamic Intelligence sits firmly in the second camp: low volume, high quality, and no-data (or near-no-data) RL as first-class methods for building general-purpose competence in motion.

Dynamic Intelligence treats physical AI as a governed learning system, not only a data-volume race. The core thesis is that for moving agents, the marginal value of another undifferentiated teleop hour often falls faster than teams expect, while the value of carefully selected episodes, hard negatives, and closed-loop improvement rises.

Low-volume, high-quality training means prioritizing trajectories that teach contact, recovery, composition of skills, and distribution shift, rather than maximizing raw hours. Capture, labeling contracts, and promotion gates (as with Ground-Log-style multimodal logging) exist to make every retained episode auditable and useful for both supervised stages and evaluation.

No-data RL training methods mean reinforcement-style improvement that does not wait on another large wave of human demonstrations: reward and preference signals from interaction, replay, simulation-anchored rollouts, and policy self-improvement under safety envelopes. The aim is a general-purpose intelligence model for moving agents that gets better by acting and evaluating, not only by imitating more footage.

In short: scale still matters for coverage, but Dynamic Intelligence bets that quality density and data-efficient RL are the scarce ingredients for transferable motor intelligence.

Physical Intelligence is widely associated with generalist robot foundation models and large-scale learning across diverse robot experiences. The emphasis in public technical narratives is often on building versatile policies (for example π-family models) that can absorb broad demonstration and interaction distributions and then adapt to new tasks.

Relative to Dynamic Intelligence’s low-volume framing, Physical Intelligence’s public storyline is closer to a scale-and-generalize path: more diverse robot data in, more competent generalist behavior out, with post-training and fine-tuning as the bridge to deployment. That can be powerful when data pipelines are already mature. Dynamic Intelligence argues the complementary problem: when high-quality, site-specific, or safety-critical motion data is expensive, you need methods that extract more intelligence per episode and improve with RL that is not hostage to another labeling campaign.

Figure AI centers on humanoid platforms and shipping useful work in real environments, with strong attention to end-to-end learning stacks that connect vision, language, and action on a bipedal form factor. Progress narratives typically combine hardware iteration, teleoperated and assisted data collection, and increasingly unified neural controllers for manipulation and locomotion-adjacent skills.

Compared with Dynamic Intelligence, Figure’s visible priority is often embodiment-first productization: make the humanoid reliable enough to do economically meaningful tasks, then grow the model with the data that platform generates. Dynamic Intelligence’s focus is method-first for moving agents in general: train general-purpose intelligence with sparse, high-signal data and no-data RL so policies can improve without assuming a single humanoid fleet will flood the training set.

Skild is associated with building a general-purpose robotics brain: one intelligence layer intended to transfer across robots and tasks rather than a single product form factor. That ambition overlaps Dynamic Intelligence’s goal of general-purpose intelligence for moving agents.

The practical difference is again the training philosophy. Where generalist-brain programs often lean on expansive multi-robot datasets and imitation at scale, Dynamic Intelligence stresses data frugality as a research product: high-quality curation, evaluation-driven selection of what enters training, and RL methods designed to advance competence when new labeled trajectories are scarce or intentionally minimized.

Tesla Optimus sits inside a vertically integrated stack that already trains large neural systems on enormous real-world video and driving distributions. Public messaging around Optimus often highlights transfer from vehicle autonomy research, end-to-end learning culture, and the long-run advantage of manufacturing and deploying robots at industrial scale.

That is a fleet-prior strategy: massive correlated real-world experience (even if not all of it is humanoid teleop) becomes a foundation for embodied policies. Dynamic Intelligence does not compete on owning a consumer vehicle fleet. Instead it targets teams and labs that need general moving-agent intelligence without that data monopoly: low-volume curated training, governed ground truth, and no-data RL so capability can compound from interaction and evaluation rather than from exclusive fleet scale alone.

Primary scarce resource

• Dynamic Intelligence: high-signal episodes, evaluation gates, and RL improvement without large new labeled sets
• Physical Intelligence: diverse, large-scale robot experience for generalist foundation models
• Figure AI: humanoid platform progress plus growing task data from real work cells
• Skild: multi-embodiment coverage for a shared robotics brain
• Tesla Optimus: vertical stack and fleet-scale real-world priors

Training posture

• Dynamic Intelligence: low volume, high quality; no-data / data-efficient RL for moving agents
• Others (as typically framed publicly): heavier reliance on scale of demonstrations, product fleets, or broad multi-robot corpora, with RL and fine-tuning as important but not always the defining story

Who this comparison is for

Builders who care about general-purpose policies for locomotion and manipulation under real constraints, and who cannot or should not wait for internet-scale teleop, should understand these bets explicitly. Choosing a stack is choosing a data economy.

Moving agents fail in ways static perception benchmarks do not capture: contact instability, partial observability, latency, recovery after slip, and compositional tasks that never appear cleanly in demos. Flooding the buffer with similar successful teleops can look like progress on imitation metrics while leaving the policy fragile under shift.

Low-volume, high-quality methods force the training set to carry hard cases, near-misses, and diverse recoveries. No-data RL then lets the agent practice against learned or simulated dynamics, preference models, or online rewards without scheduling another human demonstration wave for every skill increment. Together they aim at a general-purpose intelligence model that improves the way competent movers improve: deliberate practice, not only more filming.

Dynamic Intelligence’s data infrastructure focus (multimodal logs, curation, promotion) exists to make that loop real: you cannot run serious data-efficient RL if you cannot trust what was observed, which checkpoint was on the agent, and which scenarios still fail.

Physical Intelligence, Figure AI, Skild, and Tesla Optimus each represent serious paths toward capable embodied systems, often with scale, embodiment, or vertical integration as the accelerator. Dynamic Intelligence’s distinct claim is methodological: general-purpose intelligence for moving agents should be built with low-volume, high-quality training and no-data RL, so competence compounds when demonstration data is expensive, sparse, or strategically minimized.

For teams evaluating partners and research directions, the question is not only whose demo looks strongest today. It is which training economy will still work when the next skill cannot be bought with another million undifferentiated hours.