Skip to content
Bud's Wiki

Dyna Robotics announces Dyna-2.1 and Taku humanoid — from tasks to whole workflows

TL;DR

Section titled “TL;DR”

Dyna Robotics announces Dyna-2.1 and Taku, a new humanoid platform positioned as a step from single-task mastery to end-to-end workflow automation (“agentic whole employees”). The framing argument is a compounding-error point: even at 99% per-step success, whole-workflow success (e.g. a laundry cycle) drops below 50%, and non-linear task chaining exceeds what current single-task policies can handle. To close that gap the company redesigned the full stack — hardware (Taku), whole-body controller, and reasoning/action models. Taku is designed from first principles for data transferability and robot learning, uses large-scale GPU-parallelized RL in Isaac Sim for a robust whole-body loco-manipulation controller, and is meant to transfer physical skills from human video via the Dyna-2 world-action model pretrained on millions of hours of human video. The company claims 30 minutes to a couple of hours of on-robot data lets Taku learn new whole-body tasks with vastly different motion primitives, and the stated next milestone is deployment-flywheel operation at customer sites.

Key claims

Section titled “Key claims”
  • Real-world workflow automation is the target, not individual task mastery: at 99% per-step success, a full laundry cycle falls below 50% success — the argument for a workflow-native system rather than task-native policies [thread §2].
  • Taku is positioned as a first-principles hardware redesign inspired by Japanese takumi (匠), explicitly optimized for data transferability and robot learning as first-order design constraints [thread §4].
  • Taku’s whole-body controller is trained with large-scale GPU-parallelized RL in Isaac Sim, targeting precise loco-manipulation [thread §5].
  • Taku is designed to transfer physical skills from human videos via Dyna-2, the world-action model pretrained on millions of hours of human video (linking to Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models); with 30 min – a few hours of on-robot data, Taku is claimed to complete new whole-body tasks with very different motion primitives [thread §6–§7].
  • Stated next milestone: bring the stack to real customer sites and run a deployment flywheel where every piece of deployment data improves the whole system [thread §8].

Method

Section titled “Method”

Announcement thread with no paper or technical report attached. Concrete components named:

  • Taku — new proprietary humanoid, co-designed with the learning stack; specific DoF, sensing, and BOM not disclosed.
  • Whole-body controller — large-scale GPU-parallelized RL in Isaac Sim for precise loco-manipulation. Sibling training pattern to standard sim-to-real humanoid controller pipelines; no specifics on reward, curriculum, or transfer method.
  • Dyna-2 world-action model — the WAM pretrained on 1M+ h of human video (see Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models and infra companion Training Dyna-2 at million-hour scale, repeatably) is the pretraining substrate that Taku consumes to bootstrap new skills.
  • Post-training — 30 min to a couple of hours of on-robot data per new whole-body task, consistent with the Dyna-2 paper’s post-training regime.

No architecture diagrams, ablations, per-task success numbers, video benchmarks, or comparisons to prior Dyna-1 / Dyna-2 baselines are in the thread. Everything is at the launch-announcement level.

Results

Section titled “Results”

None disclosed. The thread makes qualitative claims — “vastly different motion primitives” learnable from ~30 min – few hours of on-robot data — with accompanying video attachments (not analyzed here). No customer-site deployment numbers, no head-to-head against Dyna-2’s own 46% → 87% joint quality/throughput/reliability pass rate from the Dyna-2 report.

Why it’s interesting

Section titled “Why it’s interesting”

This is the hardware + product-framing update that follows the pretraining-scale bet Dyna Robotics made public in Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models (1M+ hours of human video pretraining, cross-embodiment scaling law from human → robot). The Dyna-2 paper argued that world modeling + human-video scale is the substrate for embodied policies; Dyna-2.1 + Taku is the operationalization on a new humanoid designed for that substrate — the co-design here is between the robot hardware and the pretraining substrate, sibling to Getting a Grip on Robotic Data Collection — Co-Designed Koala Gripper for Force-Mediated Manipulation‘s Koala co-design (collection tool + gripper) but pushed one level higher to the full humanoid target embodiment. The framing shift — from “commercial-grade mastery on individual tasks” to “agentic whole employees” — is also directly comparable to Delta Intelligence launches Δ₀ — whole-body humanoid model with latent world-action brain, 69-DoF controller, and 10,000+ h egocentric panoramic pretraining Δ₀’s whole-body humanoid launch and GEN-1.5: Embodied Foundation Models are One-Shot Learners GEN-1.5’s one-shot generalist framing, all three converging on humanoid + human-video-pretraining + workflow-scale evaluation within weeks of each other. Distinct from Δ₀’s in-house 69-DoF controller by explicitly staying open on the “physical skills transfer from human video” framing that Dyna-2 quantified as a cross-embodiment scaling law rather than an emergent property.

See also

Section titled “See also”