DONIC Demonstrates Generalist Dexterous Bimanual Tracking Control Using Human Demonstrations
| Source: HF Papers | Original article
Researchers introduce DONIC, a generalist system that learns dexterous bimanual tracking control from human demonstrations, tackling the lack of contact‑rich bimanual data.
A new robot‑control model called **DONIC** has been released, promising to close a long‑standing gap in dexterous manipulation research. Unlike earlier systems that concentrate on single‑hand tasks, DONIC is a “generalist” bimanual tracker that learns directly from heterogeneous human demonstrations and can coordinate two robot hands to follow arbitrary reference trajectories.
The breakthrough hinges on two advances. First, the policy predicts bounded corrective actions for each hand using proprioceptive feedback, object‑state information and the target reference, allowing it to stay “reference‑conditioned” without being locked to a single demonstration. Second, the team behind DONIC assembled **BiHOI**, a new multi‑view dataset comprising 1,575 sequences and 229 objects, in which 65.9 % of frames feature sustained two‑hand contact—far above the 38.6 % ceiling of prior datasets. DONIC was trained jointly on BiHOI and five existing public hand‑object collections, using per‑hand object‑centered rewards and clip‑centered advantages to encourage coordinated contact.
Why it matters is clear: real‑world tasks—from assembling furniture to handling kitchenware—often require simultaneous use of both hands. Existing single‑hand trackers struggle with the combinatorial complexity of coordinating contacts, limiting their applicability outside controlled labs. By demonstrating that a single policy can generalise to unseen trajectories and objects without fine‑tuning, DONIC opens a path toward more versatile service and manufacturing robots.
The next steps will reveal how quickly the community adopts the model and dataset. Benchmarks on standard manipulation suites, integration with robot platforms such as Dream‑Labs’ Dream‑1 foundation model, and extensions to richer sensory modalities are likely to follow. Watch for upcoming evaluations that compare DONIC’s performance against task‑specific baselines and for any open‑source releases that could accelerate research into truly bimanual, contact‑rich robot skills.
Sources
Back to AIPULSEN