# Others stack sensors, this one tears them all apart Others stack sensors, this one tears them all apart Full article: https://haiguangboy.com/posts/blind_dexterity_whole_body_humanoid_manipulation_via_pure_proprioception ## Core method Under compliant control, joint deviation is itself an implicit tactile channel, but to see this signal, you first have to feed your own issued commands back to yourself ↳ Under compliant PD control, the deviation between commanded joint angles and measured joint angles is proportional to torque, so any motion exposes contact information in the encoder readings. But to reconstruct this deviation, the policy must include the previous step's commanded action in its observation—a design choice that already exists in much legged robot reinforcement learning but whose value is rarely made explicit ↳ On the Unitree G1, without cameras, markers, force/torque sensors, or tactile arrays, they trained four types of tasks: IMU-free push-resistant walking, stopping a soccer ball with the foot, dynamically stepping onto a skateboard, and grabbing a handle to lift a suitcase ↳ The blind policy exhibits genuine active interactive perception: lifting a foot to scan for a target, touching the skateboard's nose/tail to improve position estimation, dragging the skateboard to improve orientation estimation—not passively waiting for contact, but actively creating informative contact ## Key results · This paper's own method (no privileged information at any point, directly training a blind policy from scratch with reinforcement learning) achieved high success rates on all three tasks: soccer 92.9%, skateboard 88.2%, suitcase 83-90% · Counterexample: if you first train a "privileged teacher" that can see the object's ground-truth state, then distill a blind student, the results are far worse (31.2%, 8.2%, 34.6%)—distillation learns the teacher's "go straight for the goal" actions, and cannot learn the active search behavior the teacher never demonstrated · Feeding estimator predictions back into the observation yields no stable benefit; short-term proprioceptive history alone is usually sufficient; but estimator error still systematically decreases after contact, proving the information does exist · Letting the policy adjust joint stiffness itself stably reduces the suitcase tipping rate by 4-5 percentage points ## Related line comparison - Opposite line · [N0-Foundation: Opening a new era of touch](https://haiguangboy.com/posts/mathcaln_0_foundation_towards_the_age_of_tactile_intelligence) `mathcaln_0_foundation_towards_the_age_of_tactile_intelligence_2026_09`: N0-Foundation: Opening a new era of touch - Same line · [PAVE: Failure data can also feed a policy](https://haiguangboy.com/posts/pave_predictive_alignment_and_value_guided_evolution_for_world_action_policies) `pave_predictive_alignment_and_value_guided_evolution_for_world_action_policies_2026_09`: PAVE: Failure data can also feed a policy - Same line · [PRISM: Can scale replace explicit modeling?](https://haiguangboy.com/posts/prism_precision_and_contact_rich_real_world_industrial_skill_dataset_with_multim) `prism_precision_and_contact_rich_real_world_industrial_skill_dataset_with_multim_2026_08`: PRISM: Can scale replace explicit modeling? - Same line · [Fingertip anesthesia analogy: The missing touch in teleoperation](https://haiguangboy.com/posts/the_missing_touch_spatially_distributed_tactile_feedback_brings_teleoperation_cl) `the_missing_touch_spatially_distributed_tactile_feedback_brings_teleoperation_cl_2026_08`: Fingertip anesthesia analogy: The missing touch in teleoperation - Same line · [The more accurate the future prediction, the more sluggish the robot becomes](https://haiguangboy.com/posts/omega-0) `omega_0_a_latent_predictive_world_action_model_for_concurrent_humanoid_loco_mani_2026_08`: Real-robot results are only qualitative demonstrations, not reliability evaluations collected under a fixed protocol—all quantitative numbers come from simulation, and the authors explicitly say the real-robot experiments demonstrate transferability, not reliability estimates validates all three generalization tests—across objects, across scenes, and human data transfer—are qualitative descriptions, with no quantitative success rates given - Same line · [The robot GPT-3 moment: Plenty of story, not enough evidence](https://haiguangboy.com/posts/gen-1-5-one-shot) `generalist_blog_gen_1_5_20260818_2026_08`: Real-robot results are only qualitative demonstrations, not reliability evaluations collected under a fixed protocol—all quantitative numbers come from simulation, and the authors explicitly say the real-robot experiments demonstrate transferability, not reliability estimates validates three zero-shot generalization demonstrations: combining two physical prompts, sim-to-real transfer, and human-hand-to-robot-hand cross-embodiment imitation ## Boundaries · This implicit tactile channel has extremely low spatial resolution, defined at the joint level rather than the contact point, so precise localization is inherently harder than with dense tactile skin · Each state estimator targets only the specific object family used in training, and no category-level generalization capability is claimed · Real-robot results are only qualitative demonstrations, and all quantitative numbers come from simulation—the authors themselves say this proves transferability, not reliability ## Author's judgment (not part of the paper, a cross-paper synthesis view) This paper goes head-to-head with the already-interpreted N0-Foundation: N0-Foundation's core judgment is that pure vision + proprioception can neither expose contact states nor provide signals for mid-execution correction, so force/touch must be explicitly modeled; this paper instead proves that as long as the observation includes the previous step's command, proprioception itself can serve as a sufficient implicit tactile channel, completing tasks like ball interception and skateboard pushing that rely heavily on contact feedback, without any dedicated tactile hardware. Which side holds up better may depend on how much precision the task requires from contact information—this paper itself admits this channel has very low resolution, which happens to echo the finding of the already-interpreted Touch paper: whether tactile resolution is worth anything depends on how redundant it is with other feedback channels.