{"id":"d68b389d-026f-462b-8b74-3451fa337c67","slug":"datafarm","name":"DATAFARM","description":"DATAFARM (Distribution-Aligned Task and Motion Planning for Fine-Tuning Vision-Language-Action Models) is a 2026 research method from Stanford/Princeton (Sahoo, Huang, Silver; under review) that aligns planner-generated TAMP trajectories with a VLA's pretraining distribution in joint space, motion style, and timing so synthetic demos help fine-tuning. Project page reports 56.7% average success vs 8.3% raw TAMP and 61.7% human teleop on three tabletop tasks.","brainType":"research-model","isOpen":false,"maturityStage":"research","architecture":"TAMP trajectory generation guided by pretraining-distribution models: Gaussian mixture over joint configurations + trajectory encoder costs for endpoint selection, path optimization, and retiming; evaluated fine-tuning pi0.5-DROID-class VLAs.","reviewStatus":"reviewed","sources":[{"url":"https://prpl-group.com/datafarm/","title":"DATAFARM: Distribution-Aligned Task and Motion Planning for Fine-Tuning Vision-Language-Action Models","sourceName":"PRPL / DATAFARM project page","publishedAt":"2026"},{"url":"https://x.com/robotsdigest/status/2106429690659631214","title":"Robots Digest: DATAFARM summary","sourceName":"X (@robotsdigest)","publishedAt":"2026-10-03"}],"keyFacts":[{"label":"Average success 56.7% vs raw TAMP 8.3% / human teleop ~61.7% (research-stated)","value":"Research-stated (research-stated; also Robots Digest X): DATAFARM (Sahoo / Huang / Silver; Stanford + Princeton; under review 2026) aligns TAMP synthetic trajectories to the VLA pretraining distribution (joint configs, motion style, timing). On three tabletop fine-tuning tasks, average success 56.7% vs 8.3% raw TAMP and 61.7% human teleoperation; approaches teleop without teleoperating. Cap: author project-page table; not an independent replication."},{"label":"Timing alignment mattered most in ablations (research-stated)","value":"Research-stated (research-stated): removing one alignment component at a time drops average success from 56.7% to 23.3% without joint space, 15.0% without style, and 0% without timing. Cap: author ablation table on the project page."},{"label":"OOD cloth-folding retention ~85% after fine-tune (research-stated)","value":"Research-stated (research-stated): on Deformable Object Manipulation (cloth folding; outside fine-tuning distribution / beyond TAMP), the fine-tuned model retains about 85% success vs ~90% for the pretrained pi0.5-DROID baseline. Cap: author OOD retention claim."}],"aliases":["DATAFARM","Distribution-Aligned TAMP","DATAFARM VLA"],"collisionRisk":"low","reviewNote":null,"builtOnBrainId":null,"createdAt":"2026-10-03T21:00:56.741Z","updatedAt":"2026-10-10T21:05:54.325Z","jsonLd":{"@context":"https://schema.org","@type":"SoftwareApplication","@id":"https://registry.deploy.report/brains/datafarm","url":"https://registry.deploy.report/brains/datafarm","name":"DATAFARM","alternateName":["DATAFARM","Distribution-Aligned TAMP","DATAFARM VLA"],"description":"DATAFARM (Distribution-Aligned Task and Motion Planning for Fine-Tuning Vision-Language-Action Models) is a 2026 research method from Stanford/Princeton (Sahoo, Huang, Silver; under review) that aligns planner-generated TAMP trajectories with a VLA's pretraining distribution in joint space, motion style, and timing so synthetic demos help fine-tuning. Project page reports 56.7% average success vs 8.3% raw TAMP and 61.7% human teleop on three tabletop tasks.","identifier":"d68b389d-026f-462b-8b74-3451fa337c67","applicationCategory":"research-model","publisher":{"@id":"https://deploy.report/#organization"}},"framework_metadata":{"framework_schema_version":"0.1.0","verification_status":"verified","maturity_stage":"research","lifecycle_state":null,"architectural_position":{"cohort":null,"sub_cohorts":[]},"within_cohort_verified_vs_claimed_pair":null,"cap_flags":[],"verification_depth":{"sources_count":2,"primary_source_types":[]}}}